kayg/calternal
Watch
1
0
Fork
You've already forked calternal
 0
Code Issues 803 Pull requests Projects Releases Packages Wiki Activity Actions 430

PERF: tab switch to fully painted on a hard disk with heavy data (Calendar much slower than Photos); stress, measure, fix #549

New issue
Open
opened 2026-09-30 18:01:25 +00:00 by kayg · 65 comments
kayg commented 2026-09-30 18:01:25 +00:00
Owner
Copy link

Owner request (2026-09-30, night)

While recording a demo, the owner saw that opening the Calendar tab takes much longer than Photos. Photos had 1 photo; Calendar had 2–3 days of events. "It's very important that they load very, very, very fast", and it must be tested on a hard disk, not on fast flash storage: HDD latency and IOPS are the constraint, and if it is fast there, it is lightning fast on SSD. Stress test it: switch between every tab with lots of items loaded into each tab, measure the time until the tab is fully painted, and minimise that time. Fix layout bugs that add to the load time.

1. Environment (prove it is a hard disk)

  • Perf VM root@10.69.69.63 (flock /root/perf.lock for every run; 4 vCPU, 7 GB). Its disks are QEMU virtual disks, so the ROTA flag is not proof.
  • Before anything else, run fio 4 KiB random reads with --direct=1 on the candidate disk. Report IOPS and p50/p99 latency. An HDD shows about 100–250 IOPS and 4–15 ms. If no disk on the perf VM behaves like an HDD, stop and report; do not measure on flash.
  • Put the calternal userdata (the Home) and the SQLite databases on that disk.
  • Use the shared release build (/mnt/hdd/targets/release-shared, web build from the same commit); no local builds on the perf VM.
  • Cold runs: sync; echo 3 > /proc/sys/vm/drop_caches before each cold run, and restart the server for a cold server cache. Warm runs: no drop.

2. Data (real-shaped fixtures, tests only)

Load each tab heavily and reproducibly (a seeded generator committed under bench/):

  • Calendar: 3 years of log entries, events and tasks (e.g. 30/day), with areas, recurring events, attachments;
  • Photos: 50k photos with thumbnails;
  • Files: 100k files in nested folders;
  • Notes: 5k notes;
  • Mail: 50k messages in 3 accounts;
  • Contacts: 5k;
  • Money: 3 budgets and 20k transactions;
  • Tasks: 2k;
  • Analytics over that data.

3. Measure

Playwright on a production build, Chromium and WebKit, at 1440 and 390 px. For every ordered pair of tabs (A→B), cold and warm, record 11 interleaved runs:

  • tab switch to first paint, and to fully painted: the tab's content is visible, no skeletons or spinners, images in the viewport decoded, no layout shift after that (PerformanceObserver: LCP-equivalent via element timing on the tab's main content, layout-shift entries, long tasks);
  • requests on the critical path (count, bytes, server time from Server-Timing);
  • server-side disk reads per request (/proc/<pid>/io read_bytes delta; iostat await);
  • main-thread long tasks and INP for the tab click.
    Also the first load of each tab after sign-in. Report p50/p95 per tab.
    Budgets (decision rule, fixed up front):
  • warm tab switch to fully painted ≤ 100 ms p95;
  • cold (HDD, dropped caches) ≤ 400 ms p95 for the first viewport;
  • no CLS after first paint;
  • no request waterfall deeper than 2 on the critical path.

4. Fix

Find the causes and fix them, one commit per cause, with before and after numbers:

  • Calendar first: why is it much slower than Photos?
  • Look for N+1 requests, per-day fetches, loading whole ranges instead of the viewport, SQLite queries without indexes, reading Markdown files at request time on HDD (random I/O), synchronous parsing on the main thread, a missing cache-first render (#instant-loading rule: render from cache and revalidate), layout thrash, and unvirtualised lists.
  • Keep visual flair: no removed animations. Pointer animations run after content paints; keyboard switches never animate (#527).
  • Look for layout bugs that cost time (forced reflow loops, measuring before fonts load, resize-observer storms, CLS from late-loading chrome) and fix them.
    Server changes: per-crate Rust gates. Web: bun run check, bun run test.

5. Keep it measured

  • Commit the harness as bench/tab-switch.mjs + a bench/run.sh profile, and add the results to docs/perf/ as the new baseline, so the weekly perf timer tracks tab switching from now on.
  • Post a table (tab × cold/warm × p50/p95, before/after) on this issue.
    Rules: aggregates only; no owner data; no pushes or deploys. If a fix needs a design decision (e.g. prefetching all tabs in the background), list it under "Owner decisions" with numbers; do not build it.
## Owner request (2026-09-30, night) While recording a demo, the owner saw that opening the **Calendar** tab takes much longer than **Photos**. Photos had 1 photo; Calendar had 2–3 days of events. "It's very important that they load very, very, very fast", and it must be tested **on a hard disk, not on fast flash storage**: HDD latency and IOPS are the constraint, and if it is fast there, it is lightning fast on SSD. Stress test it: switch between every tab with **lots of items loaded into each tab**, measure the time until the tab is **fully painted**, and minimise that time. Fix layout bugs that add to the load time. ### 1. Environment (prove it is a hard disk) - Perf VM `root@10.69.69.63` (`flock /root/perf.lock` for every run; 4 vCPU, 7 GB). Its disks are QEMU virtual disks, so the ROTA flag is not proof. - Before anything else, run `fio` 4 KiB random reads with `--direct=1` on the candidate disk. Report IOPS and p50/p99 latency. An HDD shows about 100–250 IOPS and 4–15 ms. If no disk on the perf VM behaves like an HDD, stop and report; do not measure on flash. - Put the calternal userdata (the Home) and the SQLite databases on that disk. - Use the shared release build (`/mnt/hdd/targets/release-shared`, web build from the same commit); no local builds on the perf VM. - **Cold runs:** `sync; echo 3 > /proc/sys/vm/drop_caches` before each cold run, and restart the server for a cold server cache. **Warm runs:** no drop. ### 2. Data (real-shaped fixtures, tests only) Load each tab heavily and reproducibly (a seeded generator committed under `bench/`): - Calendar: 3 years of log entries, events and tasks (e.g. 30/day), with areas, recurring events, attachments; - Photos: 50k photos with thumbnails; - Files: 100k files in nested folders; - Notes: 5k notes; - Mail: 50k messages in 3 accounts; - Contacts: 5k; - Money: 3 budgets and 20k transactions; - Tasks: 2k; - Analytics over that data. ### 3. Measure Playwright on a production build, Chromium and WebKit, at 1440 and 390 px. For every ordered pair of tabs (A→B), cold and warm, record 11 interleaved runs: - **tab switch to first paint**, and **to fully painted**: the tab's content is visible, no skeletons or spinners, images in the viewport decoded, no layout shift after that (PerformanceObserver: LCP-equivalent via element timing on the tab's main content, `layout-shift` entries, long tasks); - requests on the critical path (count, bytes, server time from `Server-Timing`); - server-side disk reads per request (`/proc/<pid>/io` read_bytes delta; `iostat` await); - main-thread long tasks and INP for the tab click. Also the first load of each tab after sign-in. Report p50/p95 per tab. **Budgets (decision rule, fixed up front):** - warm tab switch to fully painted ≤ 100 ms p95; - cold (HDD, dropped caches) ≤ 400 ms p95 for the first viewport; - no CLS after first paint; - no request waterfall deeper than 2 on the critical path. ### 4. Fix Find the causes and fix them, one commit per cause, with before and after numbers: - Calendar first: why is it much slower than Photos? - Look for N+1 requests, per-day fetches, loading whole ranges instead of the viewport, SQLite queries without indexes, reading Markdown files at request time on HDD (random I/O), synchronous parsing on the main thread, a missing cache-first render (#instant-loading rule: render from cache and revalidate), layout thrash, and unvirtualised lists. - Keep visual flair: no removed animations. Pointer animations run after content paints; keyboard switches never animate (#527). - Look for layout bugs that cost time (forced reflow loops, measuring before fonts load, resize-observer storms, CLS from late-loading chrome) and fix them. Server changes: per-crate Rust gates. Web: `bun run check`, `bun run test`. ### 5. Keep it measured - Commit the harness as `bench/tab-switch.mjs` + a `bench/run.sh` profile, and add the results to `docs/perf/` as the new baseline, so the weekly perf timer tracks tab switching from now on. - Post a table (tab × cold/warm × p50/p95, before/after) on this issue. Rules: aggregates only; no owner data; no pushes or deploys. If a fix needs a design decision (e.g. prefetching all tabs in the background), list it under "Owner decisions" with numbers; do not build it.
tabswitch-549-smoke-desktop-light.zip
3.4 MiB
tab-switch-549-after-screenshots.zip
4.2 MiB
tabswitch-549-traces-and-rust-gates.zip
403 KiB
tabswitch-549-production-review.zip
8.9 MiB
tabswitch-549-final-performance.zip
266 KiB
round4-production-review.zip
8.9 MiB
round4-restored-production-review.zip
8.9 MiB
tabswitch-549-round4-final-review.zip
16 MiB
tabswitch-549-round4-full-e2e-proof.zip
786 B
kayg commented 2026-09-30 20:32:00 +00:00
Author
Owner
Copy link

Starting #549 on branch job/tabswitch-549. Current HEAD/base: aa372eef6c. I am reading the required design sections and will qualify the perf VM disk before collecting timings.

Starting #549 on branch job/tabswitch-549. Current HEAD/base: aa372eef6c9312403f59a65d5b90cb3648ff52ed. I am reading the required design sections and will qualify the perf VM disk before collecting timings.
kayg commented 2026-09-30 20:32:34 +00:00
Author
Owner
Copy link

Perf VM qualification finding: /mnt/hdd is a 100 GiB ext4 mount on /dev/sda (ROTA=1), but that flag is not sufficient proof. The required fio probe did not run because failed; I have collected no performance timings. I am making fio available, then I will run the 4 KiB direct random-read qualification under /root/perf.lock.

Perf VM qualification finding: /mnt/hdd is a 100 GiB ext4 mount on /dev/sda (ROTA=1), but that flag is not sufficient proof. The required fio probe did not run because failed; I have collected no performance timings. I am making fio available, then I will run the 4 KiB direct random-read qualification under /root/perf.lock.
kayg commented 2026-09-30 20:37:32 +00:00
Author
Owner
Copy link

Stopped at the HDD qualification gate in #549. The job requires an HDD-like disk and explicitly says to stop if no perf VM disk behaves like one.

Both disks were tested read-only with fio 3.39, 4 KiB random reads, direct I/O, libaio, queue depth 1, 60 seconds, while holding flock /root/perf.lock. Load average was captured inside each lock.

Device Role 4 KiB random-read IOPS p50 completion latency p99 completion latency Load average
/dev/sda /mnt/hdd, ext4, 100 GiB, ROTA=1 11,899 0.061 ms 0.183 ms 0.20 2.00 1.67
/dev/sdb system disk, 100 GiB, ROTA=0 7,568 0.089 ms 0.387 ms 0.06 1.40 1.48

The requested HDD behavior is about 100–250 IOPS and 4–15 ms. Neither device qualifies. ROTA was not used as proof; the direct-I/O results establish that both candidates are much faster than the required storage.

I did not place Home or SQLite data on these disks, run the tab-switch workload, create the benchmark harness, modify source files, or run build/test gates. No before/after timing exists.

Head SHA: aa372eef6c (unchanged).
Gate output: not run; the explicit HDD qualification condition stopped the job before implementation.
Known gap: tab-switch root causes, fixes, harness, baseline, visual evidence, and gates remain unaddressed until an HDD-like perf disk is available.
Owner decisions: none made.

Stopped at the HDD qualification gate in #549. The job requires an HDD-like disk and explicitly says to stop if no perf VM disk behaves like one. Both disks were tested read-only with fio 3.39, 4 KiB random reads, direct I/O, libaio, queue depth 1, 60 seconds, while holding `flock /root/perf.lock`. Load average was captured inside each lock. | Device | Role | 4 KiB random-read IOPS | p50 completion latency | p99 completion latency | Load average | |---|---|---:|---:|---:|---| | /dev/sda | /mnt/hdd, ext4, 100 GiB, ROTA=1 | 11,899 | 0.061 ms | 0.183 ms | 0.20 2.00 1.67 | | /dev/sdb | system disk, 100 GiB, ROTA=0 | 7,568 | 0.089 ms | 0.387 ms | 0.06 1.40 1.48 | The requested HDD behavior is about 100–250 IOPS and 4–15 ms. Neither device qualifies. ROTA was not used as proof; the direct-I/O results establish that both candidates are much faster than the required storage. I did not place Home or SQLite data on these disks, run the tab-switch workload, create the benchmark harness, modify source files, or run build/test gates. No before/after timing exists. Head SHA: aa372eef6c9312403f59a65d5b90cb3648ff52ed (unchanged). Gate output: not run; the explicit HDD qualification condition stopped the job before implementation. Known gap: tab-switch root causes, fixes, harness, baseline, visual evidence, and gates remain unaddressed until an HDD-like perf disk is available. Owner decisions: none made.
kayg commented 2026-09-30 20:57:47 +00:00
Author
Owner
Copy link

Starting Round 2 on job/tabswitch-549. Worktree HEAD: aa372eef6c9312403f59a65d5b90cb3648ff52ed; current origin/dev: 15e17aeafc8ea160c109e62fba57f6961c39d21f. I will first capture the flash-like host result under /root/perf.lock, then build and qualify the file-backed delayed HDD device before moving Home and SQLite data or measuring tab switches.

Starting Round 2 on `job/tabswitch-549`. Worktree HEAD: `aa372eef6c9312403f59a65d5b90cb3648ff52ed`; current `origin/dev`: `15e17aeafc8ea160c109e62fba57f6961c39d21f`. I will first capture the flash-like host result under `/root/perf.lock`, then build and qualify the file-backed delayed HDD device before moving Home and SQLite data or measuring tab switches.
kayg commented 2026-09-30 21:05:08 +00:00
Author
Owner
Copy link

HDD-emulation qualification (all runs held /root/perf.lock): the real /mnt/hdd backing device (/dev/sda, ext4) is flash-like: 4 KiB direct random read at QD1 measured 36,164 IOPS, p50 21.4 µs, p99 91.6 µs. dm-delay loaded successfully. I created a 40 GiB file-backed loop device with direct I/O, device-mapper adds 8 ms read and write delay, and ext4 is mounted at /srv/hdd-emu. The systemd-run scope reports io.max for the loop backing device as rbps=157286400 wbps=157286400 riops=200 wiops=200 (load average 0.16 before qualification).

4 KiB direct random read fio, in that scope: QD1 = 125.0 IOPS, 8.03 ms p50 / 8.16 ms p99; QD16 = 200.9 IOPS, 100.14 ms p50 / 104.33 ms p99. QD16 saturates the required 200 IOPS ceiling and shows queueing latency. I will run the server and workload together in the same limited scope for the app measurements.

HDD-emulation qualification (all runs held `/root/perf.lock`): the real `/mnt/hdd` backing device (`/dev/sda`, ext4) is flash-like: 4 KiB direct random read at QD1 measured 36,164 IOPS, p50 21.4 µs, p99 91.6 µs. `dm-delay` loaded successfully. I created a 40 GiB file-backed loop device with direct I/O, device-mapper adds 8 ms read and write delay, and ext4 is mounted at `/srv/hdd-emu`. The `systemd-run` scope reports `io.max` for the loop backing device as `rbps=157286400 wbps=157286400 riops=200 wiops=200` (load average 0.16 before qualification). 4 KiB direct random read fio, in that scope: QD1 = 125.0 IOPS, 8.03 ms p50 / 8.16 ms p99; QD16 = 200.9 IOPS, 100.14 ms p50 / 104.33 ms p99. QD16 saturates the required 200 IOPS ceiling and shows queueing latency. I will run the server and workload together in the same limited scope for the app measurements.
kayg commented 2026-09-30 23:06:49 +00:00
Author
Owner
Copy link

Smoke found two harness defects. First, CLI parsing read query before its const initialization and exited with a temporal-dead-zone error; fixed in commit d1f3d3da6. Second, the first real same-cgroup QD1 fio result had 7.963 ms p50 with the configured 8 ms delay, so the strict lower bound rejected a valid emulated-HDD result. The script accepted its 100–250 IOPS check, then failed only on p50. Commit 4a092f093 allows a 7.5–10.5 ms qualification band and ensures cleanup runs after setup or qualification failures. The production smoke is running again.

Smoke found two harness defects. First, CLI parsing read `query` before its `const` initialization and exited with a temporal-dead-zone error; fixed in commit `d1f3d3da6`. Second, the first real same-cgroup QD1 fio result had 7.963 ms p50 with the configured 8 ms delay, so the strict lower bound rejected a valid emulated-HDD result. The script accepted its 100–250 IOPS check, then failed only on p50. Commit `4a092f093` allows a 7.5–10.5 ms qualification band and ensures cleanup runs after setup or qualification failures. The production smoke is running again.
kayg commented 2026-09-30 23:11:32 +00:00
Author
Owner
Copy link

The next production smoke reached server startup and failed with AddrInUse because the shared VM already had a listener on 127.0.0.1:8080. The harness used the same fixed port for every job. Commit e4cdb7186 now probes a free app/CalDAV port pair while holding /root/perf.lock, then uses that pair for the server and both SSH forwards. Cleanup also tears down the local tunnel after this startup failure. I am rerunning the smoke with the isolated ports.

The next production smoke reached server startup and failed with `AddrInUse` because the shared VM already had a listener on `127.0.0.1:8080`. The harness used the same fixed port for every job. Commit `e4cdb7186` now probes a free app/CalDAV port pair while holding `/root/perf.lock`, then uses that pair for the server and both SSH forwards. Cleanup also tears down the local tunnel after this startup failure. I am rerunning the smoke with the isolated ports.
kayg commented 2026-09-30 23:20:40 +00:00
Author
Owner
Copy link

The first production smoke confirms the Calendar/Photos difference with fixture data on the emulated HDD. At Chromium 1440 px, light theme, one direct post-sign-in visit: Calendar fully painted in 19,472.7 ms and Photos in 5,076.7 ms. These are one-run diagnostic values, not p95 results. The same smoke's fio probe in the server cgroup measured QD1 125 IOPS (p50 8.028 ms, p99 8.159 ms) and QD16 200.919 IOPS (p50 100.139 ms, p99 104.333 ms). All seven requested tabs are enabled, and the interleaved switch matrix is still running.

The first production smoke confirms the Calendar/Photos difference with fixture data on the emulated HDD. At Chromium 1440 px, light theme, one direct post-sign-in visit: Calendar fully painted in 19,472.7 ms and Photos in 5,076.7 ms. These are one-run diagnostic values, not p95 results. The same smoke's fio probe in the server cgroup measured QD1 125 IOPS (p50 8.028 ms, p99 8.159 ms) and QD16 200.919 IOPS (p50 100.139 ms, p99 104.333 ms). All seven requested tabs are enabled, and the interleaved switch matrix is still running.
kayg commented 2026-09-30 23:38:31 +00:00
Author
Owner
Copy link

The smoke fixture generator reported 366 daily notes and 10,980 log entries for the 2026-09-30 anchor, but the JSON report initially overwrote those counts with an assumed 365/10,950 because this range includes an extra day. Commit 584c39ef2 now preserves counts from the generated fixture, and it checkpoints after every five switch samples so a time-boxed run retains partial aggregates. The active smoke was launched before that commit; I will correct its aggregate fixture metadata before using it as evidence.

The smoke fixture generator reported 366 daily notes and 10,980 log entries for the 2026-09-30 anchor, but the JSON report initially overwrote those counts with an assumed 365/10,950 because this range includes an extra day. Commit `584c39ef2` now preserves counts from the generated fixture, and it checkpoints after every five switch samples so a time-boxed run retains partial aggregates. The active smoke was launched before that commit; I will correct its aggregate fixture metadata before using it as evidence.
kayg commented 2026-10-01 00:03:00 +00:00
Author
Owner
Copy link

The seven-Tab smoke recorded first-load aggregates and the same-cgroup fio qualification, but it did not record a switch sample. Its next /root/perf.lock acquisition waited 15 minutes and exited with Error: perf lock exited early: while another perf run held the single-tenant VM. I stopped this run and confirmed the benchmark server PID had exited. No switch timing is reported from that attempt. The output still records the qualified device and first-load timings.

The seven-Tab smoke recorded first-load aggregates and the same-cgroup fio qualification, but it did not record a switch sample. Its next `/root/perf.lock` acquisition waited 15 minutes and exited with `Error: perf lock exited early:` while another perf run held the single-tenant VM. I stopped this run and confirmed the benchmark server PID had exited. No switch timing is reported from that attempt. The output still records the qualified device and first-load timings.
kayg commented 2026-10-01 00:05:29 +00:00
Author
Owner
Copy link

Source inspection found a Calendar HDD hot path: Today loads a 14-day Agenda page, then loadAgenda calls readDay() once for every day with Logs to fetch repair details. Each readDay() is a separate /api/v1/notes/journal/{date} request and reads that Daily note. AgendaList renders those repair details only when the matching day is shown. I will move those reads to the existing viewport boundary so the initial Calendar view reads only nearby days and loads later repair details as the User scrolls. This is a code-path finding; the perf lock timeout prevented a switch-request count.

Source inspection found a Calendar HDD hot path: Today loads a 14-day Agenda page, then `loadAgenda` calls `readDay()` once for every day with Logs to fetch repair details. Each `readDay()` is a separate `/api/v1/notes/journal/{date}` request and reads that Daily note. `AgendaList` renders those repair details only when the matching day is shown. I will move those reads to the existing viewport boundary so the initial Calendar view reads only nearby days and loads later repair details as the User scrolls. This is a code-path finding; the perf lock timeout prevented a switch-request count.
kayg commented 2026-10-01 01:12:38 +00:00
Author
Owner
Copy link

Post-change smoke finding: the harness currently treats a 120 s page-wide network-idle timeout as a completed paint. The 390 px first visits for Calendar and Photos hit that exact cap (120,801 ms and 120,601 ms), while other widths completed much sooner. These rows are timeouts, not paint timings. I am preserving them as invalid diagnostics and will correct the paint boundary to use visible tab content, busy markers, decoded viewport images and paint frames, then record any rerun separately. The 1440 px dark profile is still in progress; its cold server restart is running under the HDD cgroup.

Post-change smoke finding: the harness currently treats a 120 s page-wide network-idle timeout as a completed paint. The 390 px first visits for Calendar and Photos hit that exact cap (120,801 ms and 120,601 ms), while other widths completed much sooner. These rows are timeouts, not paint timings. I am preserving them as invalid diagnostics and will correct the paint boundary to use visible tab content, busy markers, decoded viewport images and paint frames, then record any rerun separately. The 1440 px dark profile is still in progress; its cold server restart is running under the HDD cgroup.
kayg commented 2026-10-01 01:18:27 +00:00
Author
Owner
Copy link

The first after-change smoke stopped during the 1440 px dark profile: a cold server restart did not reach /readyz before the harness start deadline, so that profile has no switch rows. The server log showed SQLx pool acquisition delays up to 14.56 s and a job update taking 2.81 s. The preceding 1440 px dark first-load samples recorded 1-minute load averages of 5.84 and 6.09. I stopped the attempt and verified that its PID file did not refer to a live tabswitch server. This is a host-load-limited smoke result; I have not treated it as a product crash.

The first after-change smoke stopped during the 1440 px dark profile: a cold server restart did not reach `/readyz` before the harness start deadline, so that profile has no switch rows. The server log showed SQLx pool acquisition delays up to 14.56 s and a job update taking 2.81 s. The preceding 1440 px dark first-load samples recorded 1-minute load averages of 5.84 and 6.09. I stopped the attempt and verified that its PID file did not refer to a live tabswitch server. This is a host-load-limited smoke result; I have not treated it as a product crash.
kayg commented 2026-10-01 01:18:54 +00:00
Author
Owner
Copy link

I corrected the #549 harness and committed it as c9868f93b: the full-paint boundary now waits for route content, visible busy markers to clear, viewport images to decode and two paint frames. It no longer waits for page-wide network idle or silently accepts the 120 s deadline as a completed paint. The corrected Chromium smoke is now running with the same HDD image and cgroup.

I corrected the #549 harness and committed it as `c9868f93b`: the full-paint boundary now waits for route content, visible busy markers to clear, viewport images to decode and two paint frames. It no longer waits for page-wide network idle or silently accepts the 120 s deadline as a completed paint. The corrected Chromium smoke is now running with the same HDD image and cgroup.
kayg commented 2026-10-01 01:33:13 +00:00
Author
Owner
Copy link

The 424 came from a race in the benchmark fixture: startReverseCalDavTunnel returned after a fixed 250 ms delay without checking that the VM loopback listener could reach the local provider. I changed it to poll a TCP connection through the reverse tunnel before posting the account, committed as e2d3eea51, and passed node --check bench/tab-switch.mjs. The failed setup did not modify application code; its smoke run saved no measurement rows.

The 424 came from a race in the benchmark fixture: `startReverseCalDavTunnel` returned after a fixed 250 ms delay without checking that the VM loopback listener could reach the local provider. I changed it to poll a TCP connection through the reverse tunnel before posting the account, committed as `e2d3eea51`, and passed `node --check bench/tab-switch.mjs`. The failed setup did not modify application code; its smoke run saved no measurement rows.
kayg commented 2026-10-01 01:47:35 +00:00
Author
Owner
Copy link

The targeted paint pass timed out after 120 s at the helper's busy-marker wait. The selector included #route-content itself, which is always visible, so the predicate could never succeed. I removed the root and aria-busy alone from the visual loading test; cache-first content may stay aria-busy while it revalidates. The correction is 82bcc9ddd, and node --check bench/tab-switch.mjs passes. No app loading state caused this timeout.

The targeted paint pass timed out after 120 s at the helper's busy-marker wait. The selector included `#route-content` itself, which is always visible, so the predicate could never succeed. I removed the root and `aria-busy` alone from the visual loading test; cache-first content may stay aria-busy while it revalidates. The correction is `82bcc9ddd`, and `node --check bench/tab-switch.mjs` passes. No app loading state caused this timeout.
kayg commented 2026-10-01 02:08:58 +00:00
Author
Owner
Copy link

Completed the #549 job on job/tabswitch-549.

Head: 234f046864017ac60364183cd6d8cb8fbfb90e77

Built

  • Added a 40 GB file-backed HDD emulator: O_DIRECT loop, dm-delay at 8 ms read/write, ext4, and a systemd scope limited to 200 read IOPS, 200 write IOPS, and 150 MiB/s in each direction. Setup and teardown live in bench/hdd-emu.sh.
  • Added the seeded Home generator, CalDAV fixture support, Chromium harness, and bench/run.sh --measure-only --tab-switch profile.
  • Calendar now reads repair details only for the anchor day or a day with Logs, and only when that Agenda section approaches the viewport. The existing repair eligibility rule stays in place.
  • Fixed the harness paint boundary and CalDAV tunnel readiness checks. It waits for visible route content, explicit visible loading UI, decoded viewport images, and two paint frames. It does not wait for unrelated page-wide network idle.

HDD qualification

Queue depth Read IOPS p50 p99 Load average
1 125.0 8.028 ms 8.159 ms 0.65 / 1.31 / 1.98
16 200.9 100.139 ms 104.333 ms 0.67 / 1.27 / 1.95

Queue depth 1 confirms the configured delay. Queue depth 16 reached the 200 IOPS limit and queued requests.

Corrected smoke results

This is one Chromium run per pair, on the seeded smoke fixture, in light mode. Since n=1, p50 and p95 are equal and are not percentile estimates. “Before” switch values were not measured.

Width State Switch Before full p50/p95 After first p50/p95 After full p50/p95 CLS samples
390 px Warm Photos → Calendar Not measured 9,239.5 / 9,239.5 ms 9,319 / 9,319 ms 0
390 px Cold Photos → Calendar Not measured 2,436.7 / 2,436.7 ms 12,348.7 / 12,348.7 ms 1
390 px Cold Calendar → Photos Not measured 2,405.2 / 2,405.2 ms 2,770.6 / 2,770.6 ms 0
390 px Warm Calendar → Photos Not measured 742.3 / 742.3 ms 861.4 / 861.4 ms 0
1440 px Warm Photos → Calendar Not measured 5,095.6 / 5,095.6 ms 5,219.4 / 5,219.4 ms 0
1440 px Cold Photos → Calendar Not measured 4,217.9 / 4,217.9 ms 6,490.2 / 6,490.2 ms 0
1440 px Cold Calendar → Photos Not measured 3,347.9 / 3,347.9 ms 3,870.6 / 3,870.6 ms 0
1440 px Warm Calendar → Photos Not measured 2,756.8 / 2,756.8 ms 3,721.9 / 3,721.9 ms 1

None of these rows met the fixed budgets: warm full paint was over 100 ms, and cold first paint was over 400 ms. The deepest critical path was 12 requests. The longest observed main-thread task was 5,907 ms, and click INP reached 2,304 ms. Server-Timing headers were absent; iostat await values were null. The 390 px first visits painted Calendar in 3,731.9 ms and Photos in 9,118.8 ms. At 1440 px, Calendar was 17,262.9 ms and Photos was 5,498.6 ms.

The earlier pre-change first-visit smoke reported 19,472.7 ms for Calendar and 5,076.7 ms for Photos at 1440 px, but used the old page-wide network-idle boundary. It is not a valid before/after comparison. The shared baseline remains unchanged.

Known gaps and decisions

  • The measured fixture had one year, 366 Daily notes, 10,980 Logs, 30 daily recurring Events, 100 Photos, 100 Files, 20 Notes, 20 Tasks, three Budgets and 100 transactions. It did not use the requested full data sizes.
  • The ordered-pair matrix covers only Calendar and Photos, Chromium, 390/1440 px, light mode, and one run. It does not cover WebKit, the other Tabs, 820 px measurements, dark-mode measurements, or 11 runs.
  • The remaining long browser tasks and high Calendar times need a browser trace and the full corpus. The Agenda renders all rows in its 14-day page; this smoke does not establish the cost of each code path.
  • Screenshots show Calendar and Photos at 390, 820 and 1440 px in light and dark mode: https://git.kayg.org/attachments/7212cc19-cb51-446e-a058-bb14350e0871
  • No design decision was needed. I did not prefetch all Tabs in the background. No product API or Rust code changed.

Gates

bun run check output:

Text sizes and UI shape values use shared role tokens.
UI transitions and animation options use shared motion tokens or documented exceptions.
Loading svelte-check in workspace: /home/kayg/Developer/calternal-wt/tabswitch-549/apps/web
Getting Svelte diagnostics...

svelte-check found 0 errors and 0 warnings

bun run test summary output:

 Test Files  140 passed (140)
      Tests  915 passed (915)
   Start at  02:26:24
   Duration  122.30s (transform 45%, environment 26%, import 17%, tests 8%, setup 4%)

node --check bench/tab-switch.mjs, bash -n bench/hdd-emu.sh, git diff --check, and JSON validation passed. cargo build --release -p calternal-server passed: Finished release profile [optimized] target(s) in 2m 07s. No Rust source changed, so Rust clippy/test gates were not run. cargo clean removed 7,653 files (2.5 GiB), and web build output was removed.

No push, merge or deploy was done.

Completed the #549 job on `job/tabswitch-549`. **Head:** `234f046864017ac60364183cd6d8cb8fbfb90e77` ## Built - Added a 40 GB file-backed HDD emulator: O_DIRECT loop, `dm-delay` at 8 ms read/write, ext4, and a systemd scope limited to 200 read IOPS, 200 write IOPS, and 150 MiB/s in each direction. Setup and teardown live in `bench/hdd-emu.sh`. - Added the seeded Home generator, CalDAV fixture support, Chromium harness, and `bench/run.sh --measure-only --tab-switch` profile. - Calendar now reads repair details only for the anchor day or a day with Logs, and only when that Agenda section approaches the viewport. The existing repair eligibility rule stays in place. - Fixed the harness paint boundary and CalDAV tunnel readiness checks. It waits for visible route content, explicit visible loading UI, decoded viewport images, and two paint frames. It does not wait for unrelated page-wide network idle. ## HDD qualification | Queue depth | Read IOPS | p50 | p99 | Load average | | ---: | ---: | ---: | ---: | --- | | 1 | 125.0 | 8.028 ms | 8.159 ms | 0.65 / 1.31 / 1.98 | | 16 | 200.9 | 100.139 ms | 104.333 ms | 0.67 / 1.27 / 1.95 | Queue depth 1 confirms the configured delay. Queue depth 16 reached the 200 IOPS limit and queued requests. ## Corrected smoke results This is one Chromium run per pair, on the seeded smoke fixture, in light mode. Since `n=1`, p50 and p95 are equal and are not percentile estimates. “Before” switch values were not measured. | Width | State | Switch | Before full p50/p95 | After first p50/p95 | After full p50/p95 | CLS samples | | ---: | --- | --- | --- | ---: | ---: | ---: | | 390 px | Warm | Photos → Calendar | Not measured | 9,239.5 / 9,239.5 ms | 9,319 / 9,319 ms | 0 | | 390 px | Cold | Photos → Calendar | Not measured | 2,436.7 / 2,436.7 ms | 12,348.7 / 12,348.7 ms | 1 | | 390 px | Cold | Calendar → Photos | Not measured | 2,405.2 / 2,405.2 ms | 2,770.6 / 2,770.6 ms | 0 | | 390 px | Warm | Calendar → Photos | Not measured | 742.3 / 742.3 ms | 861.4 / 861.4 ms | 0 | | 1440 px | Warm | Photos → Calendar | Not measured | 5,095.6 / 5,095.6 ms | 5,219.4 / 5,219.4 ms | 0 | | 1440 px | Cold | Photos → Calendar | Not measured | 4,217.9 / 4,217.9 ms | 6,490.2 / 6,490.2 ms | 0 | | 1440 px | Cold | Calendar → Photos | Not measured | 3,347.9 / 3,347.9 ms | 3,870.6 / 3,870.6 ms | 0 | | 1440 px | Warm | Calendar → Photos | Not measured | 2,756.8 / 2,756.8 ms | 3,721.9 / 3,721.9 ms | 1 | None of these rows met the fixed budgets: warm full paint was over 100 ms, and cold first paint was over 400 ms. The deepest critical path was 12 requests. The longest observed main-thread task was 5,907 ms, and click INP reached 2,304 ms. `Server-Timing` headers were absent; `iostat` await values were null. The 390 px first visits painted Calendar in 3,731.9 ms and Photos in 9,118.8 ms. At 1440 px, Calendar was 17,262.9 ms and Photos was 5,498.6 ms. The earlier pre-change first-visit smoke reported 19,472.7 ms for Calendar and 5,076.7 ms for Photos at 1440 px, but used the old page-wide network-idle boundary. It is not a valid before/after comparison. The shared baseline remains unchanged. ## Known gaps and decisions - The measured fixture had one year, 366 Daily notes, 10,980 Logs, 30 daily recurring Events, 100 Photos, 100 Files, 20 Notes, 20 Tasks, three Budgets and 100 transactions. It did not use the requested full data sizes. - The ordered-pair matrix covers only Calendar and Photos, Chromium, 390/1440 px, light mode, and one run. It does not cover WebKit, the other Tabs, 820 px measurements, dark-mode measurements, or 11 runs. - The remaining long browser tasks and high Calendar times need a browser trace and the full corpus. The Agenda renders all rows in its 14-day page; this smoke does not establish the cost of each code path. - Screenshots show Calendar and Photos at 390, 820 and 1440 px in light and dark mode: https://git.kayg.org/attachments/7212cc19-cb51-446e-a058-bb14350e0871 - No design decision was needed. I did not prefetch all Tabs in the background. No product API or Rust code changed. ## Gates `bun run check` output: ```text Text sizes and UI shape values use shared role tokens. UI transitions and animation options use shared motion tokens or documented exceptions. Loading svelte-check in workspace: /home/kayg/Developer/calternal-wt/tabswitch-549/apps/web Getting Svelte diagnostics... svelte-check found 0 errors and 0 warnings ``` `bun run test` summary output: ```text Test Files 140 passed (140) Tests 915 passed (915) Start at 02:26:24 Duration 122.30s (transform 45%, environment 26%, import 17%, tests 8%, setup 4%) ``` `node --check bench/tab-switch.mjs`, `bash -n bench/hdd-emu.sh`, `git diff --check`, and JSON validation passed. `cargo build --release -p calternal-server` passed: `Finished release profile [optimized] target(s) in 2m 07s`. No Rust source changed, so Rust clippy/test gates were not run. `cargo clean` removed 7,653 files (2.5 GiB), and web build output was removed. No push, merge or deploy was done.
kayg commented 2026-10-01 02:14:45 +00:00
Author
Owner
Copy link

Starting Round 3 on job/tabswitch-549, base/head 234f046864017ac60364183cd6d8cb8fbfb90e77. Read CLAUDE.md, CONTEXT.md and Calendar/design requirements. I will capture a Chromium 390 px warm Photos → Calendar CDP timeline + CPU profile + network log on the existing qualified HDD emulator before product changes, then post top self-time functions and the request waterfall. No push or deploy.

Starting Round 3 on `job/tabswitch-549`, base/head `234f046864017ac60364183cd6d8cb8fbfb90e77`. Read CLAUDE.md, CONTEXT.md and Calendar/design requirements. I will capture a Chromium 390 px warm Photos → Calendar CDP timeline + CPU profile + network log on the existing qualified HDD emulator before product changes, then post top self-time functions and the request waterfall. No push or deploy.
kayg commented 2026-10-01 02:24:25 +00:00
Author
Owner
Copy link

Trace attempt 1 reached the seeded fixture and Calendar screenshot, then failed before any switch: goto: Timeout 120000ms exceeded, navigating to /photos, waiting until load. No trace or switch timing is claimed. The browser runs on the shared build host (load average 29.53 / 30.81 / 29.47 at diagnosis). Commit 06cecc230 uses domcontentloaded followed by the explicit existing viewport paint boundary; off-screen image loads must not define fully painted. It also adds the CPU trace summarizer and a one-pair diagnostic option. Retrying the requested 390 px warm trace; no product code has changed.

Trace attempt 1 reached the seeded fixture and Calendar screenshot, then failed before any switch: `goto: Timeout 120000ms exceeded`, navigating to `/photos`, waiting until `load`. No trace or switch timing is claimed. The browser runs on the shared build host (load average 29.53 / 30.81 / 29.47 at diagnosis). Commit `06cecc230` uses `domcontentloaded` followed by the explicit existing viewport paint boundary; off-screen image loads must not define fully painted. It also adds the CPU trace summarizer and a one-pair diagnostic option. Retrying the requested 390 px warm trace; no product code has changed.
kayg commented 2026-10-01 02:38:06 +00:00
Author
Owner
Copy link

The second trace setup also timed out before a switch (/photos, waiting for domcontentloaded). Its Calendar screenshot contained only “Loading earlier days…”. This showed that the old fullyPainted prerequisite could accept an empty Agenda before its data arrived. Commit 778b37db6 waits while .agenda[aria-busy=true] has no .day, while still allowing cached days during revalidation. Diagnostic priming uses the Tab Bar instead of a second full-document navigation. Retrying the same seeded 390 px warm trace. Product code is unchanged.

The second trace setup also timed out before a switch (`/photos`, waiting for `domcontentloaded`). Its Calendar screenshot contained only “Loading earlier days…”. This showed that the old `fullyPainted` prerequisite could accept an empty Agenda before its data arrived. Commit `778b37db6` waits while `.agenda[aria-busy=true]` has no `.day`, while still allowing cached days during revalidation. Diagnostic priming uses the Tab Bar instead of a second full-document navigation. Retrying the same seeded 390 px warm trace. Product code is unchanged.
kayg commented 2026-10-01 02:43:32 +00:00
Author
Owner
Copy link

Round 3 trace completed before product changes, using Chromium at 390 px, light, the Round 2 smoke fixture and the existing qualified HDD cgroup. Warm Photos → Calendar: 1,781.8 ms first / 1,816.8 ms full, largest switch-time task 1,070.15 ms, CLS 0, disk-read delta 0. This is one diagnostic run, not p95. The earlier 5,907 ms task was not reproduced and cannot be attributed from this trace.

The leading app function Jr maps to packages/ui/src/time.ts:303 (systemTimeZone), called by Yr (formatClockTime) while mounting the Calendar Agenda. It constructs a fresh Intl.DateTimeFormat and calls resolvedOptions() for each row clock, even though clock label formatters are already cached. The remaining cost is Svelte DOM/style work, icon construction and GC while mounting all 14 days. wa maps to Svelte dom/elements/style.js:31; T to Lucide buildLucideIconNode.js:19; Ar to Svelte runtime.js:225.

Top 10 sampled self-time functions below use only samples after the click. The trace setup task (1,203 ms before the click) is excluded. Raw traces stay in ignored artifacts.

Function Location (zero-based line:column) Sampled self ms
Jr BCIpNbpT.js:1:70340 232.56
(garbage collector) 👎-1 98.94
getPropertyValue 👎-1 86.07
before 👎-1 65.79
setAttribute 👎-1 42.51
cloneNode 👎-1 36.57
wa Df0kZKHp.js:2:1688 27.14
query :5412:16 26.57
T DQ2PRTbu.js:0:521 22.61
Ar Df0kZKHp.js:0:22465 22.17

Longest RunTask ms: [1070.15, 313.88, 97.92, 59.48, 50.8]

Request Start ms Response ms End ms Bytes Status
/api/v1/files/events 0.0 pending pending 0
/api/v1/auth/sessions 16.54 pending pending 0
/api/v1/calendar/preferences 27.31 pending pending 0
/api/v1/auth/me 28.46 pending pending 0

The Calendar range was already cached after correct priming; no range or Tasks request was on this warm switch's data path. The four logged requests were still pending at full paint and did not block it. The trace does not reproduce the Round 2 depth-12 path. Step 2 will add server phase timing and replay request disk counters, followed by separate fixes with measurements.

Round 3 trace completed before product changes, using Chromium at 390 px, light, the Round 2 smoke fixture and the existing qualified HDD cgroup. Warm Photos → Calendar: 1,781.8 ms first / 1,816.8 ms full, largest switch-time task 1,070.15 ms, CLS 0, disk-read delta 0. This is one diagnostic run, not p95. The earlier 5,907 ms task was not reproduced and cannot be attributed from this trace. The leading app function `Jr` maps to `packages/ui/src/time.ts:303` (`systemTimeZone`), called by `Yr` (`formatClockTime`) while mounting the Calendar Agenda. It constructs a fresh `Intl.DateTimeFormat` and calls `resolvedOptions()` for each row clock, even though clock label formatters are already cached. The remaining cost is Svelte DOM/style work, icon construction and GC while mounting all 14 days. `wa` maps to Svelte `dom/elements/style.js:31`; `T` to Lucide `buildLucideIconNode.js:19`; `Ar` to Svelte `runtime.js:225`. Top 10 sampled self-time functions below use only samples after the click. The trace setup task (1,203 ms before the click) is excluded. Raw traces stay in ignored artifacts. | Function | Location (zero-based line:column) | Sampled self ms | | --- | --- | ---: | | Jr | BCIpNbpT.js:1:70340 | 232.56 | | (garbage collector) | :-1:-1 | 98.94 | | getPropertyValue | :-1:-1 | 86.07 | | before | :-1:-1 | 65.79 | | setAttribute | :-1:-1 | 42.51 | | cloneNode | :-1:-1 | 36.57 | | wa | Df0kZKHp.js:2:1688 | 27.14 | | query | :5412:16 | 26.57 | | T | DQ2PRTbu.js:0:521 | 22.61 | | Ar | Df0kZKHp.js:0:22465 | 22.17 | Longest RunTask ms: [1070.15, 313.88, 97.92, 59.48, 50.8] | Request | Start ms | Response ms | End ms | Bytes | Status | | --- | ---: | ---: | ---: | ---: | ---: | | /api/v1/files/events | 0.0 | pending | pending | 0 | | | /api/v1/auth/sessions | 16.54 | pending | pending | 0 | | | /api/v1/calendar/preferences | 27.31 | pending | pending | 0 | | | /api/v1/auth/me | 28.46 | pending | pending | 0 | | The Calendar range was already cached after correct priming; no range or Tasks request was on this warm switch's data path. The four logged requests were still pending at full paint and did not block it. The trace does not reproduce the Round 2 depth-12 path. Step 2 will add server phase timing and replay request disk counters, followed by separate fixes with measurements.
kayg commented 2026-10-01 03:16:05 +00:00
Author
Owner
Copy link

Round 3 finding: the old release server's serial startup probe returned Calendar range in 67.10 ms (258,231 bytes), Tasks/day in 116.70 ms, Files entries in 120.30 ms, but Journal/2026-10-01 took 387,005.80 ms, consumed 33,780 ms server CPU, and RSS rose from 258,207,744 to 507,092,992 bytes. All four physical read-byte deltas were zero. These are process-wide counters, including background work.

The Notes startup reconciler holds the same per-User lock that Journal GET acquires. Photos indexing readiness does not prove Notes reconciliation readiness. I am instrumenting the routes before attributing this delay to either reconciliation lock wait or Journal work. The probe now has a 30-second bound; SLOW-only startup probes are reported without stopping the Tab matrix. No Calendar cause is inferred from this isolated Journal result.

Round 3 finding: the old release server's serial startup probe returned Calendar range in 67.10 ms (258,231 bytes), Tasks/day in 116.70 ms, Files entries in 120.30 ms, but Journal/2026-10-01 took **387,005.80 ms**, consumed 33,780 ms server CPU, and RSS rose from 258,207,744 to 507,092,992 bytes. All four physical read-byte deltas were zero. These are process-wide counters, including background work. The Notes startup reconciler holds the same per-User lock that Journal GET acquires. Photos indexing readiness does not prove Notes reconciliation readiness. I am instrumenting the routes before attributing this delay to either reconciliation lock wait or Journal work. The probe now has a 30-second bound; SLOW-only startup probes are reported without stopping the Tab matrix. No Calendar cause is inferred from this isolated Journal result.
kayg commented 2026-10-01 03:43:26 +00:00
Author
Owner
Copy link

Round 3 application finding: the interrupted warm sample rendered Calendar but kept Photos selected in the Tab Bar. Evidence: artifacts/round3/local-before/chromium-390-light-failure.png.

apps/web/src/routes/+layout.svelte captures the requested pathname in pendingHeaderNavigation. Its afterNavigate callback returns unless that pathname exactly equals the completed pathname. Calendar enters at /today, which redirects to /calendar/today/<date>. A warm navigation snapshots the Photos header; the redirect then leaves that snapshot in place. Cold navigation can hide the error because its pending header already says Calendar. This explains the selected-Tab timeout; it is an application error, not a slow server probe.

The fix accepts a completed redirect within the requested Tab, then clears the navigation snapshot using the existing navigation ID guard. I will verify repeated warm Photos → Calendar returns in a production browser. The separate main-thread clock and Agenda mounting findings remain relevant to actual paint time.

Round 3 application finding: the interrupted warm sample rendered Calendar but kept Photos selected in the Tab Bar. Evidence: `artifacts/round3/local-before/chromium-390-light-failure.png`. `apps/web/src/routes/+layout.svelte` captures the requested pathname in `pendingHeaderNavigation`. Its `afterNavigate` callback returns unless that pathname exactly equals the completed pathname. Calendar enters at `/today`, which redirects to `/calendar/today/<date>`. A warm navigation snapshots the Photos header; the redirect then leaves that snapshot in place. Cold navigation can hide the error because its pending header already says Calendar. This explains the selected-Tab timeout; it is an application error, not a slow server probe. The fix accepts a completed redirect within the requested Tab, then clears the navigation snapshot using the existing navigation ID guard. I will verify repeated warm Photos → Calendar returns in a production browser. The separate main-thread clock and Agenda mounting findings remain relevant to actual paint time.
kayg commented 2026-10-01 03:49:56 +00:00
Author
Owner
Copy link

Round 3 application finding: the interrupted warm sample rendered Calendar but kept Photos selected in the Tab Bar. Evidence: artifacts/round3/local-before/chromium-390-light-failure.png.

apps/web/src/routes/+layout.svelte captures the requested pathname in pendingHeaderNavigation. Its afterNavigate callback returns unless that pathname exactly equals the completed pathname. Calendar enters at /today, which redirects to /calendar/today/<date>. A warm navigation snapshots the Photos header; the redirect then leaves that snapshot in place. Cold navigation can hide the error because its pending header already says Calendar. This explains the selected-Tab timeout; it is an application error, not a slow server probe.

The fix accepts a completed redirect within the requested Tab, then clears the navigation snapshot using the existing navigation ID guard. I will verify repeated warm Photos → Calendar returns in a production browser. The separate main-thread clock and Agenda mounting findings remain relevant to actual paint time.

The full web suite passed before the redirect fix: Test Files 140 passed (140) and Tests 918 passed (918). The redirect fix passes web checking with zero errors and warnings. The production Calendar e2e now checks three repeated Photos → Calendar returns at 390, 820 and 1440 px in light and dark, and records Agenda screenshots.

Round 3 application finding: the interrupted warm sample rendered Calendar but kept Photos selected in the Tab Bar. Evidence: `artifacts/round3/local-before/chromium-390-light-failure.png`. `apps/web/src/routes/+layout.svelte` captures the requested pathname in `pendingHeaderNavigation`. Its `afterNavigate` callback returns unless that pathname exactly equals the completed pathname. Calendar enters at `/today`, which redirects to `/calendar/today/<date>`. A warm navigation snapshots the Photos header; the redirect then leaves that snapshot in place. Cold navigation can hide the error because its pending header already says Calendar. This explains the selected-Tab timeout; it is an application error, not a slow server probe. The fix accepts a completed redirect within the requested Tab, then clears the navigation snapshot using the existing navigation ID guard. I will verify repeated warm Photos → Calendar returns in a production browser. The separate main-thread clock and Agenda mounting findings remain relevant to actual paint time. The full web suite passed before the redirect fix: `Test Files 140 passed (140)` and `Tests 918 passed (918)`. The redirect fix passes web checking with zero errors and warnings. The production Calendar e2e now checks three repeated Photos → Calendar returns at 390, 820 and 1440 px in light and dark, and records Agenda screenshots.
kayg commented 2026-10-01 04:02:06 +00:00
Author
Owner
Copy link

Round 3 cache finding: the saved local warm Photos → Calendar sample after interleaved process restarts transferred 267,688 bytes in four requests, took 8,668 ms, and had a 3,534 ms main-thread task. The process-wide server counter also included 4,150 ms CPU and 180,269,056 physical read bytes from startup background work.

subscribeCalendarFileChanges erased Calendar navigation caches on each Files notification. Its coalesced page refresh erased them again and emptied the Agenda before awaiting replacement rows. This can turn a previously visited Calendar into a data-cold return during startup reconciliation.

Calendar now opts into retaining bounded stale snapshots, rejects pre-change pending cache fills, paints a cached Agenda synchronously, and revalidates the visible page. Quiet Agenda refresh keeps its rows until replacement data arrives. Auth changes still erase cached data. Legacy callers retain the old clear behavior; no existing test expectation was changed. Two new tests prove stale paint during a delayed refresh and rejection of an older pending cache fill. The targeted 15 tests passed and web checking reported zero errors and warnings.

Round 3 cache finding: the saved local warm Photos → Calendar sample after interleaved process restarts transferred 267,688 bytes in four requests, took 8,668 ms, and had a 3,534 ms main-thread task. The process-wide server counter also included 4,150 ms CPU and 180,269,056 physical read bytes from startup background work. `subscribeCalendarFileChanges` erased Calendar navigation caches on each Files notification. Its coalesced page refresh erased them again and emptied the Agenda before awaiting replacement rows. This can turn a previously visited Calendar into a data-cold return during startup reconciliation. Calendar now opts into retaining bounded stale snapshots, rejects pre-change pending cache fills, paints a cached Agenda synchronously, and revalidates the visible page. Quiet Agenda refresh keeps its rows until replacement data arrives. Auth changes still erase cached data. Legacy callers retain the old clear behavior; no existing test expectation was changed. Two new tests prove stale paint during a delayed refresh and rejection of an older pending cache fill. The targeted 15 tests passed and web checking reported zero errors and warnings.
kayg commented 2026-10-01 04:37:51 +00:00
Author
Owner
Copy link

Round 3 update: the instrumented local cold trace is saved at artifacts/round3/local-cold-trace/. This is a local fallback with a frozen production SPA; it is not an HDD measurement. The HDD attempt exited before lock acquisition. The cold trace used the clock, deferred-day and redirect fixes, before the two-day range change. One sample: first marker 631.8 ms, full paint 709.7 ms; longest traced task 235.74 ms. The old automation marker can include a loading shell, so it does not verify the cold first-viewport budget. The harness now records a separate input-to-Calendar-content marker inside the browser, while retaining the old metric for comparison.

Selected server phase probe, same fixture (366 Daily notes, 10,980 Logs): Calendar 15-day response: 361.03 ms wall, total;dur=333.439, db;dur=14.100, parse;dur=2.587, 257,397 bytes, 0 physical read bytes. Tasks: 12.39 ms wall, total;dur=3.426, db;dur=3.364, parse;dur=0.000, 0 reads. Files: 33.97 ms wall, total;dur=15.195, db;dur=0.382, parse;dur=0.000, 0 reads. Journal exceeded the bounded 30 s timeout; process read delta 20,480 bytes, CPU delta 9,150 ms, RSS 495,828,992 bytes. Process counters include startup background work. These are not exclusive request costs. Timing phases cover selected operations; total includes uninstrumented work.

Decision: Today starts with two days, matching the immediate Agenda mount. Older pages remain fourteen days. The preloader uses the same initial range key. A write refresh retains all loaded days. The real production e2e checks that the initial API request covers two days plus the one-day Log carry-in. No new endpoint is needed: the range endpoint already exists.

The required origin/dev merge is being resolved with both sides retained. Web check: svelte-check found 0 errors and 0 warnings. Web tests: Test Files 140 passed (140) and Tests 937 passed (937). Final Rust gates are running.

Function Location (zero-based line:column) Sampled self ms
getPropertyValue 👎-1 17.33
getBoundingClientRect 👎-1 14.59
Le kqO9BAQK.js:102:12998 14.07
anonymous 👎-1 13.76
appendChild 👎-1 12.13
isElementHiddenForAria :3545:31 11.73
get Df0kZKHp.js:2:10682 10.36
(garbage collector) 👎-1 8.74
reducedMotion kqO9BAQK.js:110:142524 7.22
i 0.BxpVL5jn.js:5:73090 6.92

Longest RunTask ms: [235.74, 50.79]

Request Start ms Response ms End ms Bytes Status
/api/v1/files/thumb/8e4b12f3a0881aceae771d159784710eac10125aaa2805ad52fd1bc81cefba12 0.0 270.58 270.76 57 404
/api/v1/tags 194.75 514.58 507.72 774 200
/api/v1/calendar/tags 195.58 302.14 278.95 791 200
/_app/immutable/nodes/54.Cat-UBBS.js 230.64 286.86 256.64 377 200
/_app/immutable/chunks/BelWyZaW.js 236.89 299.15 257.13 242 200
/api/v1/files/events 278.41 pending pending 0
/api/v1/calendar/range 314.01 382.66 807.43 266927 200
/_app/immutable/nodes/16.DIhrYNgy.js 331.59 380.49 359.83 32588 200
/_app/immutable/chunks/FY6ssK1K.js 334.05 380.15 361.65 5471 200
/_app/immutable/assets/16.BszGowaM.css 339.25 369.06 363.55 2691 200
Round 3 update: the instrumented local cold trace is saved at `artifacts/round3/local-cold-trace/`. This is a local fallback with a frozen production SPA; it is not an HDD measurement. The HDD attempt exited before lock acquisition. The cold trace used the clock, deferred-day and redirect fixes, before the two-day range change. One sample: first marker 631.8 ms, full paint 709.7 ms; longest traced task 235.74 ms. The old automation marker can include a loading shell, so it does not verify the cold first-viewport budget. The harness now records a separate input-to-Calendar-content marker inside the browser, while retaining the old metric for comparison. Selected server phase probe, same fixture (366 Daily notes, 10,980 Logs): Calendar 15-day response: 361.03 ms wall, `total;dur=333.439, db;dur=14.100, parse;dur=2.587`, 257,397 bytes, 0 physical read bytes. Tasks: 12.39 ms wall, `total;dur=3.426, db;dur=3.364, parse;dur=0.000`, 0 reads. Files: 33.97 ms wall, `total;dur=15.195, db;dur=0.382, parse;dur=0.000`, 0 reads. Journal exceeded the bounded 30 s timeout; process read delta 20,480 bytes, CPU delta 9,150 ms, RSS 495,828,992 bytes. Process counters include startup background work. These are not exclusive request costs. Timing phases cover selected operations; total includes uninstrumented work. Decision: Today starts with two days, matching the immediate Agenda mount. Older pages remain fourteen days. The preloader uses the same initial range key. A write refresh retains all loaded days. The real production e2e checks that the initial API request covers two days plus the one-day Log carry-in. No new endpoint is needed: the range endpoint already exists. The required origin/dev merge is being resolved with both sides retained. Web check: `svelte-check found 0 errors and 0 warnings`. Web tests: `Test Files 140 passed (140)` and `Tests 937 passed (937)`. Final Rust gates are running. | Function | Location (zero-based line:column) | Sampled self ms | | --- | --- | ---: | | getPropertyValue | :-1:-1 | 17.33 | | getBoundingClientRect | :-1:-1 | 14.59 | | Le | kqO9BAQK.js:102:12998 | 14.07 | | anonymous | :-1:-1 | 13.76 | | appendChild | :-1:-1 | 12.13 | | isElementHiddenForAria | :3545:31 | 11.73 | | get | Df0kZKHp.js:2:10682 | 10.36 | | (garbage collector) | :-1:-1 | 8.74 | | reducedMotion | kqO9BAQK.js:110:142524 | 7.22 | | i | 0.BxpVL5jn.js:5:73090 | 6.92 | Longest RunTask ms: [235.74, 50.79] | Request | Start ms | Response ms | End ms | Bytes | Status | | --- | ---: | ---: | ---: | ---: | ---: | | /api/v1/files/thumb/8e4b12f3a0881aceae771d159784710eac10125aaa2805ad52fd1bc81cefba12 | 0.0 | 270.58 | 270.76 | 57 | 404 | | /api/v1/tags | 194.75 | 514.58 | 507.72 | 774 | 200 | | /api/v1/calendar/tags | 195.58 | 302.14 | 278.95 | 791 | 200 | | /_app/immutable/nodes/54.Cat-UBBS.js | 230.64 | 286.86 | 256.64 | 377 | 200 | | /_app/immutable/chunks/BelWyZaW.js | 236.89 | 299.15 | 257.13 | 242 | 200 | | /api/v1/files/events | 278.41 | pending | pending | 0 | | | /api/v1/calendar/range | 314.01 | 382.66 | 807.43 | 266927 | 200 | | /_app/immutable/nodes/16.DIhrYNgy.js | 331.59 | 380.49 | 359.83 | 32588 | 200 | | /_app/immutable/chunks/FY6ssK1K.js | 334.05 | 380.15 | 361.65 | 5471 | 200 | | /_app/immutable/assets/16.BszGowaM.css | 339.25 | 369.06 | 363.55 | 2691 | 200 |
kayg commented 2026-10-01 05:23:37 +00:00
Author
Owner
Copy link

The local 11-run Photos → Calendar matrix completed (88 samples): Chromium and WebKit, warm and cold, 390 and 1440 px, light mode. This matrix used a frozen production SPA at the shared-window stage, before the later request-zone and paging fixes. The server was frozen as well. It is a local fallback, not HDD evidence. The unmodified old build stopped after one sample on the selected-header bug; the intermediate three-fix build stopped after four samples on a cold missing-Tab failure. The old build does not provide eleven valid before samples.

Engine Width State n Full p50/p95 ms Post-paint shift samples
chromium 390 warm 11 1388.9/3577.3 2
chromium 390 cold 11 1717.2/3568.7 7
chromium 1440 warm 11 1890.3/6888.9 6
chromium 1440 cold 11 2961.9/6177.6 10
webkit 390 warm 11 2154/5339 0
webkit 390 cold 11 2793/3659 0
webkit 1440 warm 11 5604/9305 0
webkit 1440 cold 11 3664/5515 0

WebKit does not expose the layout-shift observer in this run: its zero shift counts do not prove zero CLS. The historic full marker did not require TimeGrid placement and date chunks; the next harness revision does, so cold desktop values are diagnostic rather than verified first-viewport times. No budget has passed.

New evidence: several phone shifts identify BUTTON.status.more-days at y=511, then outside the viewport. The shared sentinel observer immediately requested another fourteen days when the initial two-day page was short. This defeated the range bound and moved the visible footer. The fix requires the end to approach from outside the viewport before automatic paging; the manual action remains. The paging control also keeps one node and reserves both labels while loading. New tests preserve paging after a real approach and verify the retained node.

The Calendar adapter also resolved the device zone per Event and converted two midnight bounds per Event. It now captures the query zone for async projection, computes bounds once per day and reuses the existing formatters. The new 120-Event test verifies two zone lookups and the query zone after preferences change in flight. No existing test expectation was changed.

The full Calendar e2e exposed a setup bug: its N helper was called while the Composer field was still open, typing a literal n into the frozen snapshot. Closing the verified empty Composer before the next flow preserves all old assertions. A later original pointer click was blocked by the toast list rectangle, outside the visible toast. The fix lets empty list space pass input through while visible toasts keep pointer handling.

All five Rust crate gates passed after origin/dev merge; the latest extra timing slice also passed Files, Calendar and Server gates. The updated full web gate and Calendar e2e are running.

The local 11-run Photos → Calendar matrix completed (88 samples): Chromium and WebKit, warm and cold, 390 and 1440 px, light mode. This matrix used a frozen production SPA at the shared-window stage, before the later request-zone and paging fixes. The server was frozen as well. It is a local fallback, not HDD evidence. The unmodified old build stopped after one sample on the selected-header bug; the intermediate three-fix build stopped after four samples on a cold missing-Tab failure. The old build does not provide eleven valid before samples. | Engine | Width | State | n | Full p50/p95 ms | Post-paint shift samples | | --- | ---: | --- | ---: | ---: | ---: | | chromium | 390 | warm | 11 | 1388.9/3577.3 | 2 | | chromium | 390 | cold | 11 | 1717.2/3568.7 | 7 | | chromium | 1440 | warm | 11 | 1890.3/6888.9 | 6 | | chromium | 1440 | cold | 11 | 2961.9/6177.6 | 10 | | webkit | 390 | warm | 11 | 2154/5339 | 0 | | webkit | 390 | cold | 11 | 2793/3659 | 0 | | webkit | 1440 | warm | 11 | 5604/9305 | 0 | | webkit | 1440 | cold | 11 | 3664/5515 | 0 | WebKit does not expose the layout-shift observer in this run: its zero shift counts do not prove zero CLS. The historic full marker did not require TimeGrid placement and date chunks; the next harness revision does, so cold desktop values are diagnostic rather than verified first-viewport times. No budget has passed. New evidence: several phone shifts identify `BUTTON.status.more-days` at y=511, then outside the viewport. The shared sentinel observer immediately requested another fourteen days when the initial two-day page was short. This defeated the range bound and moved the visible footer. The fix requires the end to approach from outside the viewport before automatic paging; the manual action remains. The paging control also keeps one node and reserves both labels while loading. New tests preserve paging after a real approach and verify the retained node. The Calendar adapter also resolved the device zone per Event and converted two midnight bounds per Event. It now captures the query zone for async projection, computes bounds once per day and reuses the existing formatters. The new 120-Event test verifies two zone lookups and the query zone after preferences change in flight. No existing test expectation was changed. The full Calendar e2e exposed a setup bug: its N helper was called while the Composer field was still open, typing a literal n into the frozen snapshot. Closing the verified empty Composer before the next flow preserves all old assertions. A later original pointer click was blocked by the toast list rectangle, outside the visible toast. The fix lets empty list space pass input through while visible toasts keep pointer handling. All five Rust crate gates passed after origin/dev merge; the latest extra timing slice also passed Files, Calendar and Server gates. The updated full web gate and Calendar e2e are running.
kayg commented 2026-10-01 05:51:46 +00:00
Author
Owner
Copy link

Round 3 HDD measurement is running with the lock. Qualification: QD1 125 IOPS, p50 8.028 ms; QD16 200.946 IOPS, p50 100.139 ms. The current 11-run matrix freezes server/frontend at 67db40161.

Per-request process-I/O probe: Calendar range 200, wall 2056.453 ms, total;dur=11.073, db;dur=1.666, fs;dur=0.041, parse;dur=5.963, physical read_bytes 0; Files entries 200, wall1198.118 ms, total;dur=1034.562, db;dur=1033.788, parse;dur=0.000, read_bytes0; Journal exceeded the 30s probe limit during reconciliation, read_bytes0. Process counters include background work. Wall includes the browser/build-host transport and is not Server-Timing total.

The new warm390 trace has longest RunTask626.011ms. Its largest sampled self-time is getPropertyValue227.49ms, called by chunk CDN-Za5M.js function e. That chunk is apps/web/src/lib/ui/uiScale.svelte.ts: its root MutationObserver calls getComputedStyle for every root style/class mutation, even when --ui-scale does not change. Current tokens only set scale on :root, the coarse-pointer media query, and explicit inline scale overrides; theme classes do not set it. I will filter observer work to actual inline --ui-scale changes and retain pointer-change reads. This preserves CSS as the token source.

Calendar e2e also exposed a real same-Log refresh race: ItemPreview resets editing on any item object change; LogEntryEditor then resets text from its refreshed log prop. A stable Log identity plus an edit-session snapshot now has a passing regression test (14 Agenda tests). Existing e2e assertions are unchanged.

Round 3 HDD measurement is running with the lock. Qualification: QD1 125 IOPS, p50 8.028 ms; QD16 200.946 IOPS, p50 100.139 ms. The current 11-run matrix freezes server/frontend at 67db40161. Per-request process-I/O probe: Calendar range 200, wall 2056.453 ms, `total;dur=11.073, db;dur=1.666, fs;dur=0.041, parse;dur=5.963`, physical read_bytes 0; Files entries 200, wall1198.118 ms, `total;dur=1034.562, db;dur=1033.788, parse;dur=0.000`, read_bytes0; Journal exceeded the 30s probe limit during reconciliation, read_bytes0. Process counters include background work. Wall includes the browser/build-host transport and is not Server-Timing total. The new warm390 trace has longest RunTask626.011ms. Its largest sampled self-time is getPropertyValue227.49ms, called by chunk CDN-Za5M.js function e. That chunk is `apps/web/src/lib/ui/uiScale.svelte.ts`: its root MutationObserver calls getComputedStyle for every root style/class mutation, even when --ui-scale does not change. Current tokens only set scale on :root, the coarse-pointer media query, and explicit inline scale overrides; theme classes do not set it. I will filter observer work to actual inline --ui-scale changes and retain pointer-change reads. This preserves CSS as the token source. Calendar e2e also exposed a real same-Log refresh race: ItemPreview resets editing on any item object change; LogEntryEditor then resets text from its refreshed log prop. A stable Log identity plus an edit-session snapshot now has a passing regression test (14 Agenda tests). Existing e2e assertions are unchanged.
kayg commented 2026-10-01 06:03:58 +00:00
Author
Owner
Copy link

Round 3 progress: atomic commits eeac28920 preserve same-Log drafts, 51032dfe1 rejects unplaced/loading TimeGrid shells in the full-paint marker, 9a4bf94e8 filters UI-scale computed-style reads, and b8ee4e580 keeps independent Composer e2e flows closed and adds production Log-editor screenshots. The existing e2e assertions are unchanged.

Web gates: svelte-check found 0 errors and 0 warnings; Test Files 141 passed (141); Tests 942 passed (942). Production build passed. Rust per-crate gates passed for calternal-plugin, Calendar, Notes, Files and calternal-server; no Rust changes since those gates.

The locked HDD matrix freezes 67db40161, before the UI-scale and draft fixes. Its completed Chromium390 group has 11 warm and 11 cold samples. Warm browser DOM marker p50/p95 1052.0/1366.6ms; historic automation-inclusive first1428.8/2985.2ms, full1515.2/3151.8ms. Cold historic first787.8/2330.6ms, full1336.0/2677.3ms. Only4 cold samples saved a DOM marker before the full marker, so no 11-sample cold DOM percentile is claimed. Warm100ms and cold400ms budgets are not met. Warm3/11 and cold4/11 recorded shifts after the historic first marker; that marker can include loading chrome. The zero-CLS budget is unverified.

The matrix is now measuring1440px. I will run the later scale fix separately with the same harness and record its exact variant. Full Calendar e2e and the single adversarial round wait for the benchmark browser, to keep one browser active at a time. I will finish the report by the5.5h cutoff even if measurements remain incomplete.

Round 3 progress: atomic commits `eeac28920` preserve same-Log drafts, `51032dfe1` rejects unplaced/loading TimeGrid shells in the full-paint marker, `9a4bf94e8` filters UI-scale computed-style reads, and `b8ee4e580` keeps independent Composer e2e flows closed and adds production Log-editor screenshots. The existing e2e assertions are unchanged. Web gates: `svelte-check found 0 errors and 0 warnings`; `Test Files 141 passed (141)`; `Tests 942 passed (942)`. Production build passed. Rust per-crate gates passed for calternal-plugin, Calendar, Notes, Files and calternal-server; no Rust changes since those gates. The locked HDD matrix freezes `67db40161`, before the UI-scale and draft fixes. Its completed Chromium390 group has 11 warm and 11 cold samples. Warm browser DOM marker p50/p95 1052.0/1366.6ms; historic automation-inclusive first1428.8/2985.2ms, full1515.2/3151.8ms. Cold historic first787.8/2330.6ms, full1336.0/2677.3ms. Only4 cold samples saved a DOM marker before the full marker, so no 11-sample cold DOM percentile is claimed. Warm100ms and cold400ms budgets are not met. Warm3/11 and cold4/11 recorded shifts after the historic first marker; that marker can include loading chrome. The zero-CLS budget is unverified. The matrix is now measuring1440px. I will run the later scale fix separately with the same harness and record its exact variant. Full Calendar e2e and the single adversarial round wait for the benchmark browser, to keep one browser active at a time. I will finish the report by the5.5h cutoff even if measurements remain incomplete.
kayg commented 2026-10-01 06:16:55 +00:00
Author
Owner
Copy link

CDP traces and full per-crate Rust gate logs are attached: tabswitch-549-traces-and-rust-gates.zip. trace3-before is the required pre-code warm390 trace. hdd-matrix-latest is the later trace at67db40161 before the UI-scale fix.

Chromium1440 completed11warm+11cold samples: at1440 warm first2504.1/4406.2ms and full2700.8/6392.0ms; cold first840.6/1289.2ms and full1074.4/1811.5ms. Only2 cold1440 samples saved the browser DOM marker, so no11-sample cold DOM percentile is claimed. Budgets remain unmet/unverified. WebKit390 is running now.

CDP traces and full per-crate Rust gate logs are attached: [tabswitch-549-traces-and-rust-gates.zip](https://git.kayg.org/attachments/0a69bb3a-d146-48ce-9bb5-d0ef0ae729c1). `trace3-before` is the required pre-code warm390 trace. `hdd-matrix-latest` is the later trace at67db40161 before the UI-scale fix. Chromium1440 completed11warm+11cold samples: at1440 warm first2504.1/4406.2ms and full2700.8/6392.0ms; cold first840.6/1289.2ms and full1074.4/1811.5ms. Only2 cold1440 samples saved the browser DOM marker, so no11-sample cold DOM percentile is claimed. Budgets remain unmet/unverified. WebKit390 is running now.
kayg commented 2026-10-01 06:54:01 +00:00
Author
Owner
Copy link

Latest full Calendar e2e at the production UI after9a4bf94e8 passes the earlier Turn-into-note step, but fails at the Daily note repair fixture (apps/web/e2e/calendar.mjs, expected201 retained): Files tus replace returns {status:412, body:{error:{code:"conflict",message:"upload destination changed"}}}. The fixture now targets the same Notes/<date>-dailynote.md that it reads, instead of creating a second file inNotes/Journal.

The412 comes from a destination guard in Files uploads (parent identity, existence or item/hash snapshot), not from a missing If-Match header: absent If-Match is allowed by the existing API. A trial Notes reconcile before reading the fixture did not resolve it and was reverted. No expectation was weakened. The exact changed fingerprint/identity has not been isolated. The full Calendar gate is failing; the header/review path is running separately for visual evidence. This is tracked here as an unresolved finding.

The complete88-sample HDD matrix is committed atb9fa4a0f6. With the remaining time, the later UI-scale comparison will prioritize11interleaved warm/cold Chromium samples at390px. Latest-code1440px andWebKit remeasurement may remain incomplete. Budgets are not met.

Latest full Calendar e2e at the production UI after9a4bf94e8 passes the earlier Turn-into-note step, but fails at the Daily note repair fixture (`apps/web/e2e/calendar.mjs`, expected201 retained): Files tus replace returns `{status:412, body:{error:{code:"conflict",message:"upload destination changed"}}}`. The fixture now targets the same `Notes/<date>-dailynote.md` that it reads, instead of creating a second file inNotes/Journal. The412 comes from a destination guard in Files uploads (parent identity, existence or item/hash snapshot), not from a missing If-Match header: absent If-Match is allowed by the existing API. A trial Notes reconcile before reading the fixture did not resolve it and was reverted. No expectation was weakened. The exact changed fingerprint/identity has not been isolated. The full Calendar gate is failing; the header/review path is running separately for visual evidence. This is tracked here as an unresolved finding. The complete88-sample HDD matrix is committed atb9fa4a0f6. With the remaining time, the later UI-scale comparison will prioritize11interleaved warm/cold Chromium samples at390px. Latest-code1440px andWebKit remeasurement may remain incomplete. Budgets are not met.
kayg commented 2026-10-01 07:30:39 +00:00
Author
Owner
Copy link

Round3 result: Calendar is still above budget. This job has substantive fixes, code and atomic commits, but it is not ready to claim the performance target or a passing full Calendar gate.

Built: cheap fixed-name Server-Timing middleware with selected DB/FS/parse/lock intervals; shared ICU formatter metadata and UTC wall-clock labels; captured request-zone/day bounds; bounded stale Calendar snapshots with auth clearing and stale-fill rejection; a two-day initial Agenda window, near-viewport day/repair mounting, deferred keyboard focus and stable paging control; redirected Tab header completion; same-Log draft preservation; UI-scale reads limited to actual scale changes; toast blank-space pointer pass-through; CDP/network/disk-read measurement and frozen HTTPS browser profiles. No new dependencies.

Trace: the pre-code warm390 trace did not reproduce the original5907 ms task. It captured1070.15 ms; top self-time232.56 ms was systemTimeZone (Jr), reached from Agenda Event clock labels. A later warm390 trace captured626.01 ms and227.49 ms in the root uiScale computed-style callback. Top10 functions and waterfalls were posted before product edits and attached. Recurrence was already server-side; the15-day range probe spent5.963 ms in selected parse work.

Measurements: complete88-sample locked HDD matrix at67db40161, before the later scale/draft fixes;11 warm+11 cold each forChromium/WebKit at390/1440 px, light, one-year10,980-Log fixture. Warm Chromium390 DOM p50/p951052.0/1366.6ms. Historic full-paint p95: Chromium390 3151.8ms, Chromium1440 6392.0ms, WebKit390 7402ms, WebKit1440 9666ms. These do not meet the warm100 ms target. Cold DOM markers are incomplete; no valid11-sample cold DOM p95 is claimed. Cold400 ms, zeroCLS and causal depth2 are not proved. Original-build11-run baseline stops at its warm header bug.

Production evidence:36screenshots coverAgenda/Day/Week/Log editor at390/820/1440 px, light/dark. Geometry checks passed; the visual reviewer remains the orchestrator. Screenshots and web/e2e/probe logs. Original/later CDP traces and full Rust logs.

Known gaps: fullCalendar e2e fails its unchanged201 assertion when the same-path Daily-note tus fixture replace returns412 upload destination changed; exact changed parent/file identity has not been isolated. Notes reconciliation still scans twice and Journal hit the30s probe limit. Plugin total excludes earlier session/access middleware and body streaming; physical I/O deltas include background work and all process file reads. WebKit observer absence is not zeroCLS/long-task evidence. Day deferral is not full row virtualization. Latest-build warm390 now measured; cold/1440/WebKit matrix, largest fixture, burst, keyboard latency and all-seven-Tab matrix are incomplete. Browser shared-host load limits wall-clock attribution. No formal baseline regression ratio is claimed.

Decisions: two initial Agenda days, fourteen-day later pages,180px deferred-day minimum from the existing intrinsic size, manual paging until the sentinel approaches from outside, a separatelock timing phase, per-request projection zone, and an edit-session Log snapshot. Theme classes do not set ui-scale; observe pointer and inline scale changes. DESIGN specifies the principles, not these bounds and field names.

Gates (verbatim excerpts; complete logs attached):

scope-clippy-plugin.log:
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 15.29s
scope-test-plugin.log:
    Finished `test` profile [unoptimized + debuginfo] target(s) in 17.49s
test result: ok. 24 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 3.45s
test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s
final-clippy-calternal-plugin-notes.log:
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 32.00s
final-test-calternal-plugin-notes.log:
    Finished `test` profile [unoptimized + debuginfo] target(s) in 1m 07s
test result: ok. 131 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 67.01s
test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.39s
test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s
extended-clippy-calternal-plugin-files.log:
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 17.42s
extended-test-calternal-plugin-files.log:
    Finished `test` profile [unoptimized + debuginfo] target(s) in 47.72s
test result: ok. 146 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 104.30s
test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s
extended-clippy-calternal-plugin-calendar.log:
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 24.84s
extended-test-calternal-plugin-calendar.log:
    Finished `test` profile [unoptimized + debuginfo] target(s) in 1m 04s
test result: ok. 80 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 3.96s
test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.12s
test result: ok. 3 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.12s
test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s
extended-clippy-calternal-server.log:
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 30.10s
extended-test-calternal-server.log:
    Finished `test` profile [unoptimized + debuginfo] target(s) in 2m 09s
test result: ok. 93 passed; 0 failed; 3 ignored; 0 measured; 0 filtered out; finished in 14.93s
web:
svelte-check found 0 errors and 0 warnings
 Test Files  141 passed (141)
      Tests  942 passed (942)
calendar header e2e: all geometry checks passed
calendar Event tag probe: Unicode/bidi, 65536-byte category, 10 malformed inputs, 24 parallel tag/range/search reads, and Calendar/Notes/Files Server-Timing passed

Full Calendar gate failure (verbatim):

AssertionError [ERR_ASSERTION]: the server accepted the Daily note fixture replacement
+ {
+   body: '{"error":{"code":"conflict","message":"upload destination changed"}}',
+   status: 412
+ }
- 201

cargo fmt --check produced no output and exited0. Web production build passed. Rust gates were per crate. The single adversarial round passed after correcting a new test setup assumption: its fixture has Events but no Daily note, so Notes success is tested through Tasks/day and missing Journal retains404. All prior expectations are unchanged.

Files:

CONTEXT.md
apps/web/e2e/animation-trace.mjs
apps/web/e2e/calendar.mjs
apps/web/e2e/files-paste.mjs
apps/web/e2e/harness.mjs
apps/web/src/lib/calendar/agenda.svelte.test.ts
apps/web/src/lib/calendar/data.test.ts
apps/web/src/lib/calendar/data.ts
apps/web/src/lib/components/AppToaster.svelte
apps/web/src/lib/navigation/modePreload.ts
apps/web/src/lib/time.test.ts
apps/web/src/lib/ui/uiScale.svelte.test.ts
apps/web/src/lib/ui/uiScale.svelte.ts
apps/web/src/routes/+layout.svelte
apps/web/src/routes/calendar/[view]/[date]/+page.svelte
bench/hdd-emu.sh
bench/run.sh
bench/tab-switch-seed.py
bench/tab-switch-trace.py
bench/tab-switch.mjs
crates/calternal-plugin/src/lib.rs
crates/calternal-plugin/src/timing.rs
crates/plugins/calendar/src/lib.rs
crates/plugins/calendar/src/view.rs
crates/plugins/files/src/lib.rs
crates/plugins/files/src/listing.rs
crates/plugins/notes/src/lib.rs
crates/plugins/notes/src/store.rs
crates/plugins/notes/src/tasks_api.rs
docs/DESIGN.md
docs/perf/2026-10-01-tabswitch-549.md
docs/perf/runs/tab-switch-2026-10-01-549-after-smoke.json
docs/perf/runs/tab-switch-2026-10-01-549-partial.json
docs/perf/runs/tab-switch-2026-10-01-549-round3-hdd.json
docs/perf/runs/tab-switch-2026-10-01-549-round3-scale-warm.json
packages/ui/src/components/calendar/AgendaList.svelte
packages/ui/src/components/calendar/ItemPreview.svelte
packages/ui/src/components/calendar/model.ts
packages/ui/src/time.ts
tests/adversarial/calendar_event_tags.mjs

HEAD: 2c10cc92f7. Required origin/dev fetch/merge done once (merge7bfd3676a). No push, deploy or merge into dev/main. Cleanup: cargo clean completed; apps/web/build and apps/web/.svelte-kit removed; no own perf servers remain; HDD emulation stays configured. Latest-build comparison: 11 warm Chromium390 px HDD samples completed on embedded build7f7b86528. DOM p50/p95623.7/768.0 ms; full933.8/1241.6 ms;7 samples with shifts after the historic marker. Latest single trace longest task152.88 ms; getAnimations48.23 ms, getBoundingClientRect31.04 ms, GC22.36 ms; root computed-style callback absent from top10. Earlier run interleaved cold switches; later run was warm-only, so this is diagnostic. The short-readiness warm/cold attempt took no samples. Final measurements and trace.

Round3 result: Calendar is still above budget. This job has substantive fixes, code and atomic commits, but it is not ready to claim the performance target or a passing full Calendar gate. Built: cheap fixed-name Server-Timing middleware with selected DB/FS/parse/lock intervals; shared ICU formatter metadata and UTC wall-clock labels; captured request-zone/day bounds; bounded stale Calendar snapshots with auth clearing and stale-fill rejection; a two-day initial Agenda window, near-viewport day/repair mounting, deferred keyboard focus and stable paging control; redirected Tab header completion; same-Log draft preservation; UI-scale reads limited to actual scale changes; toast blank-space pointer pass-through; CDP/network/disk-read measurement and frozen HTTPS browser profiles. No new dependencies. Trace: the pre-code warm390 trace did not reproduce the original5907 ms task. It captured1070.15 ms; top self-time232.56 ms was systemTimeZone (Jr), reached from Agenda Event clock labels. A later warm390 trace captured626.01 ms and227.49 ms in the root uiScale computed-style callback. Top10 functions and waterfalls were posted before product edits and attached. Recurrence was already server-side; the15-day range probe spent5.963 ms in selected parse work. Measurements: complete88-sample locked HDD matrix at67db40161, before the later scale/draft fixes;11 warm+11 cold each forChromium/WebKit at390/1440 px, light, one-year10,980-Log fixture. Warm Chromium390 DOM p50/p951052.0/1366.6ms. Historic full-paint p95: Chromium390 3151.8ms, Chromium1440 6392.0ms, WebKit390 7402ms, WebKit1440 9666ms. These do not meet the warm100 ms target. Cold DOM markers are incomplete; no valid11-sample cold DOM p95 is claimed. Cold400 ms, zeroCLS and causal depth2 are not proved. Original-build11-run baseline stops at its warm header bug. Production evidence:36screenshots coverAgenda/Day/Week/Log editor at390/820/1440 px, light/dark. Geometry checks passed; the visual reviewer remains the orchestrator. [Screenshots and web/e2e/probe logs](https://git.kayg.org/attachments/ab168123-5f2b-461d-8167-5624150ab0f5). [Original/later CDP traces and full Rust logs](https://git.kayg.org/attachments/0a69bb3a-d146-48ce-9bb5-d0ef0ae729c1). Known gaps: fullCalendar e2e fails its unchanged201 assertion when the same-path Daily-note tus fixture replace returns412 `upload destination changed`; exact changed parent/file identity has not been isolated. Notes reconciliation still scans twice and Journal hit the30s probe limit. Plugin total excludes earlier session/access middleware and body streaming; physical I/O deltas include background work and all process file reads. WebKit observer absence is not zeroCLS/long-task evidence. Day deferral is not full row virtualization. Latest-build warm390 now measured; cold/1440/WebKit matrix, largest fixture, burst, keyboard latency and all-seven-Tab matrix are incomplete. Browser shared-host load limits wall-clock attribution. No formal baseline regression ratio is claimed. Decisions: two initial Agenda days, fourteen-day later pages,180px deferred-day minimum from the existing intrinsic size, manual paging until the sentinel approaches from outside, a separatelock timing phase, per-request projection zone, and an edit-session Log snapshot. Theme classes do not set ui-scale; observe pointer and inline scale changes. DESIGN specifies the principles, not these bounds and field names. Gates (verbatim excerpts; complete logs attached): ```text scope-clippy-plugin.log: Finished `dev` profile [unoptimized + debuginfo] target(s) in 15.29s scope-test-plugin.log: Finished `test` profile [unoptimized + debuginfo] target(s) in 17.49s test result: ok. 24 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 3.45s test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s final-clippy-calternal-plugin-notes.log: Finished `dev` profile [unoptimized + debuginfo] target(s) in 32.00s final-test-calternal-plugin-notes.log: Finished `test` profile [unoptimized + debuginfo] target(s) in 1m 07s test result: ok. 131 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 67.01s test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.39s test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s extended-clippy-calternal-plugin-files.log: Finished `dev` profile [unoptimized + debuginfo] target(s) in 17.42s extended-test-calternal-plugin-files.log: Finished `test` profile [unoptimized + debuginfo] target(s) in 47.72s test result: ok. 146 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 104.30s test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s extended-clippy-calternal-plugin-calendar.log: Finished `dev` profile [unoptimized + debuginfo] target(s) in 24.84s extended-test-calternal-plugin-calendar.log: Finished `test` profile [unoptimized + debuginfo] target(s) in 1m 04s test result: ok. 80 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 3.96s test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.12s test result: ok. 3 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.12s test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s extended-clippy-calternal-server.log: Finished `dev` profile [unoptimized + debuginfo] target(s) in 30.10s extended-test-calternal-server.log: Finished `test` profile [unoptimized + debuginfo] target(s) in 2m 09s test result: ok. 93 passed; 0 failed; 3 ignored; 0 measured; 0 filtered out; finished in 14.93s web: svelte-check found 0 errors and 0 warnings Test Files 141 passed (141) Tests 942 passed (942) calendar header e2e: all geometry checks passed calendar Event tag probe: Unicode/bidi, 65536-byte category, 10 malformed inputs, 24 parallel tag/range/search reads, and Calendar/Notes/Files Server-Timing passed ``` Full Calendar gate failure (verbatim): ```text AssertionError [ERR_ASSERTION]: the server accepted the Daily note fixture replacement + { + body: '{"error":{"code":"conflict","message":"upload destination changed"}}', + status: 412 + } - 201 ``` `cargo fmt --check` produced no output and exited0. Web production build passed. Rust gates were per crate. The single adversarial round passed after correcting a new test setup assumption: its fixture has Events but no Daily note, so Notes success is tested through Tasks/day and missing Journal retains404. All prior expectations are unchanged. Files: ```text CONTEXT.md apps/web/e2e/animation-trace.mjs apps/web/e2e/calendar.mjs apps/web/e2e/files-paste.mjs apps/web/e2e/harness.mjs apps/web/src/lib/calendar/agenda.svelte.test.ts apps/web/src/lib/calendar/data.test.ts apps/web/src/lib/calendar/data.ts apps/web/src/lib/components/AppToaster.svelte apps/web/src/lib/navigation/modePreload.ts apps/web/src/lib/time.test.ts apps/web/src/lib/ui/uiScale.svelte.test.ts apps/web/src/lib/ui/uiScale.svelte.ts apps/web/src/routes/+layout.svelte apps/web/src/routes/calendar/[view]/[date]/+page.svelte bench/hdd-emu.sh bench/run.sh bench/tab-switch-seed.py bench/tab-switch-trace.py bench/tab-switch.mjs crates/calternal-plugin/src/lib.rs crates/calternal-plugin/src/timing.rs crates/plugins/calendar/src/lib.rs crates/plugins/calendar/src/view.rs crates/plugins/files/src/lib.rs crates/plugins/files/src/listing.rs crates/plugins/notes/src/lib.rs crates/plugins/notes/src/store.rs crates/plugins/notes/src/tasks_api.rs docs/DESIGN.md docs/perf/2026-10-01-tabswitch-549.md docs/perf/runs/tab-switch-2026-10-01-549-after-smoke.json docs/perf/runs/tab-switch-2026-10-01-549-partial.json docs/perf/runs/tab-switch-2026-10-01-549-round3-hdd.json docs/perf/runs/tab-switch-2026-10-01-549-round3-scale-warm.json packages/ui/src/components/calendar/AgendaList.svelte packages/ui/src/components/calendar/ItemPreview.svelte packages/ui/src/components/calendar/model.ts packages/ui/src/time.ts tests/adversarial/calendar_event_tags.mjs ``` HEAD: 2c10cc92f7ce1bf001cc4dcdad8a3c549c2e8931. Required origin/dev fetch/merge done once (merge7bfd3676a). No push, deploy or merge into dev/main. Cleanup: cargo clean completed; apps/web/build and apps/web/.svelte-kit removed; no own perf servers remain; HDD emulation stays configured. Latest-build comparison: 11 warm Chromium390 px HDD samples completed on embedded build7f7b86528. DOM p50/p95623.7/768.0 ms; full933.8/1241.6 ms;7 samples with shifts after the historic marker. Latest single trace longest task152.88 ms; getAnimations48.23 ms, getBoundingClientRect31.04 ms, GC22.36 ms; root computed-style callback absent from top10. Earlier run interleaved cold switches; later run was warm-only, so this is diagnostic. The short-readiness warm/cold attempt took no samples. [Final measurements and trace](https://git.kayg.org/attachments/90fe9d1c-bdff-4c9f-9a32-a66b92d084ea).
kayg commented 2026-10-01 08:18:50 +00:00
Author
Owner
Copy link

Round 4 starts on job/tabswitch-549, base/head 2c10cc92f7ce1bf001cc4dcdad8a3c549c2e8931. Scope: separate Journal projection reads from Notes reconciliation, render the cached Calendar snapshot before revalidation, investigate the fixture replacement 412, and repeat cold/warm browser measurements. No push or deploy.

Round 4 starts on `job/tabswitch-549`, base/head `2c10cc92f7ce1bf001cc4dcdad8a3c549c2e8931`. Scope: separate Journal projection reads from Notes reconciliation, render the cached Calendar snapshot before revalidation, investigate the fixture replacement 412, and repeat cold/warm browser measurements. No push or deploy.
kayg commented 2026-10-01 08:32:49 +00:00
Author
Owner
Copy link

Round 4 evidence: Journal GET acquires the Notes User guard and calls ensure_ids over every pending Daily note before reading one day (crates/plugins/notes/src/lib.rs). Startup/watch reconciliation holds that same guard for both Tasks and Notes scans. The new read projection commits the source with Journal resource ETags and reads through the WAL reader pool; its regression test holds both the guard and an uncommitted SQLite writer and requires the old committed day within 100 ms. Gates are running; no passing-gate claim yet.

The local production pre-change desktop trace completed: cached Photos → Calendar, full paint 3309.2 ms, one sample, largest task 429.49 ms. No Calendar range/Journal read on the switch critical path. Top sampled self cost: getBoundingClientRect 328.52 ms; TimeGrid scroll callback Re 291.60 ms; scrollTo 103.12 ms; DOM before 96.88 ms; GC 72.10 ms; getAnimations 61.86 ms. This is a shared-host local diagnostic, not an HDD percentile. The 390 px trace is running.

Two client findings are fixed in code: Tab selection waited for preloadMode (including API reads) and the Agenda first render was empty until its loading effect. Navigation now starts at selection, and initial Agenda state comes from the retained range before effects/awaits. Journal PATCH/DELETE also reused indefinitely cached ETags. They now read the current stable block immediately before If-Match; three new tests pass. The Daily-note upload 412 remains a separate Files destination identity/hash guard, not a Journal ETag header or DAV sync-token epoch; exact failed guard still under investigation.

Round 4 evidence: Journal GET acquires the Notes User guard and calls `ensure_ids` over every pending Daily note before reading one day (`crates/plugins/notes/src/lib.rs`). Startup/watch reconciliation holds that same guard for both Tasks and Notes scans. The new read projection commits the source with Journal resource ETags and reads through the WAL reader pool; its regression test holds both the guard and an uncommitted SQLite writer and requires the old committed day within 100 ms. Gates are running; no passing-gate claim yet. The local production pre-change desktop trace completed: cached Photos → Calendar, full paint 3309.2 ms, one sample, largest task 429.49 ms. No Calendar range/Journal read on the switch critical path. Top sampled self cost: `getBoundingClientRect` 328.52 ms; TimeGrid scroll callback `Re` 291.60 ms; `scrollTo` 103.12 ms; DOM `before` 96.88 ms; GC 72.10 ms; `getAnimations` 61.86 ms. This is a shared-host local diagnostic, not an HDD percentile. The 390 px trace is running. Two client findings are fixed in code: Tab selection waited for `preloadMode` (including API reads) and the Agenda first render was empty until its loading effect. Navigation now starts at selection, and initial Agenda state comes from the retained range before effects/awaits. Journal PATCH/DELETE also reused indefinitely cached ETags. They now read the current stable block immediately before If-Match; three new tests pass. The Daily-note upload 412 remains a separate Files destination identity/hash guard, not a Journal ETag header or DAV sync-token epoch; exact failed guard still under investigation.
kayg commented 2026-10-01 08:59:05 +00:00
Author
Owner
Copy link

Round 4 finding: the Daily note replacement 412 is a Files tus parent check, separate from cached Log ETags and the Notes DAV epoch. A focused real Root/tus test creates Notes/day.md, opens a replacement upload, then writes an unrelated sibling through Root before PATCH. Before the fix it returns 412 (expected 204). The request has no If-Match. The directory inode and target bytes did not change; its mtime did.

The fix uses the captured parent item ID plus device/inode for this directory destination check. It keeps pending-namespace denial, full grant checks, and target ID/hash checks. The same regression also replaces the directory inode and requires 412 with both original and replacement paths unchanged. No existing test expectation changed. This is a small Files change required by #549's Calendar fixture diagnosis.

Client commits: d91b507ce reads fresh ETags immediately before Log PATCH/DELETE; ee28ca343 mounts the cached Agenda before preload awaits; 8de003cbf tracks label transitions without getAnimations; f3b1c8f87 releases a Tab entry header in the navigation update. Notes projection commit: 799ab19ce. The legacy-import compatibility test now passes without changing its expectation. Full crate gates are in progress.

Round 4 finding: the Daily note replacement 412 is a Files tus parent check, separate from cached Log ETags and the Notes DAV epoch. A focused real Root/tus test creates `Notes/day.md`, opens a replacement upload, then writes an unrelated sibling through Root before PATCH. Before the fix it returns 412 (expected 204). The request has no If-Match. The directory inode and target bytes did not change; its mtime did. The fix uses the captured parent item ID plus device/inode for this directory destination check. It keeps pending-namespace denial, full grant checks, and target ID/hash checks. The same regression also replaces the directory inode and requires 412 with both original and replacement paths unchanged. No existing test expectation changed. This is a small Files change required by #549's Calendar fixture diagnosis. Client commits: `d91b507ce` reads fresh ETags immediately before Log PATCH/DELETE; `ee28ca343` mounts the cached Agenda before preload awaits; `8de003cbf` tracks label transitions without getAnimations; `f3b1c8f87` releases a Tab entry header in the navigation update. Notes projection commit: `799ab19ce`. The legacy-import compatibility test now passes without changing its expectation. Full crate gates are in progress.
kayg commented 2026-10-01 09:08:57 +00:00
Author
Owner
Copy link

The single requested fetch/merge is complete: merge commit 307d8221a, merging 4dc9ca394 from origin/dev. The shared remote-tracking ref then advanced while other jobs merged. The required final migration check against origin/dev (1af8ead26) found Notes migrations 21, 22 and 23 in use. Commit 8bf8c3729 renumbers the rebuildable Journal snapshot migration to 24; no deployed number is reused. I am keeping the owner's one-merge rule and am not repeatedly merging the moving ref.

Final web gates: svelte-check found 0 errors and 0 warnings; Test Files 142 passed (142); Tests 946 passed (946). Rust gates and the current release build are running with four jobs, no incremental output, and line-table debug info. The latest local trace attempt did not produce samples: the fixture restart did not reach readyz within 300 seconds under host load 36.48 and I/O pressure (full avg10 50.47%). This is recorded as failed setup, not as a latency sample. The perf VM lock was busy on the one availability check.

The single requested fetch/merge is complete: merge commit `307d8221a`, merging `4dc9ca394` from origin/dev. The shared remote-tracking ref then advanced while other jobs merged. The required final migration check against origin/dev (`1af8ead26`) found Notes migrations 21, 22 and 23 in use. Commit `8bf8c3729` renumbers the rebuildable Journal snapshot migration to 24; no deployed number is reused. I am keeping the owner's one-merge rule and am not repeatedly merging the moving ref. Final web gates: `svelte-check found 0 errors and 0 warnings`; `Test Files 142 passed (142)`; `Tests 946 passed (946)`. Rust gates and the current release build are running with four jobs, no incremental output, and line-table debug info. The latest local trace attempt did not produce samples: the fixture restart did not reach readyz within 300 seconds under host load 36.48 and I/O pressure (full avg10 50.47%). This is recorded as failed setup, not as a latency sample. The perf VM lock was busy on the one availability check.
kayg commented 2026-10-01 09:15:43 +00:00
Author
Owner
Copy link

Production UI evidence for Round 4: round4-production-review.zip (9369225 bytes). This bundle has Agenda, Day, Week and Log editor screenshots at 390, 820 and 1440 px, light and dark, plus 3x header crops and the passing geometry/redirect review log. It also includes the web gate logs and the tus regression's before/after output. Screenshots remain ignored and were not committed. The UI is the current production build; this focused visual run used the frozen existing server binary while the new server release build is in progress. The final Calendar fixture run will use the new server.

Production UI evidence for Round 4: [round4-production-review.zip](https://git.kayg.org/attachments/0930e507-9815-4203-9523-ed1bdceb1bf1) (9369225 bytes). This bundle has Agenda, Day, Week and Log editor screenshots at 390, 820 and 1440 px, light and dark, plus 3x header crops and the passing geometry/redirect review log. It also includes the web gate logs and the tus regression's before/after output. Screenshots remain ignored and were not committed. The UI is the current production build; this focused visual run used the frozen existing server binary while the new server release build is in progress. The final Calendar fixture run will use the new server.
kayg commented 2026-10-01 09:40:02 +00:00
Author
Owner
Copy link

Owner (2026-10-01): "Calendar events still take a long time to appear, by long I mean close to a second - it should be instant! It is not like they are animating either, they simply take time to load." Confirms round 4 of this issue (warm switch target ≤ 100 ms; the cached snapshot must render synchronously on tab selection).

Owner (2026-10-01): "Calendar events still take a long time to appear, by long I mean close to a second - it should be instant! It is not like they are animating either, they simply take time to load." Confirms round 4 of this issue (warm switch target ≤ 100 ms; the cached snapshot must render synchronously on tab selection).
kayg commented 2026-10-01 09:43:07 +00:00
Author
Owner
Copy link

Round 4 desktop finding: loadChunk in the Calendar route waits for Promise.all([loadRange(...), loadItems(...)]) before it fills days. The retained range alone does not contain an explicitly completed standalone rail. A warm Week/Day remount therefore still starts empty and awaits the chunk read.

The view now retains a completed chunk with its filtered rail in the existing bounded, generation-aware Calendar cache. Initial grid state reads this snapshot synchronously. The publisher captures the request's zone, photo preference and cache generation before awaiting; a completed old request cannot refill a changed or signed-out User's cache. An items error does not replace a good completed snapshot. Three new tests cover complete rail retention, photo preference separation, auth reset and invalidation races. The data and Journal targeted tests pass (24 tests). Final web gates are running for this last client slice.

Notes final gates passed:

    Finished `dev` profile [unoptimized + debuginfo] target(s) in 14m 10s
test result: ok. 132 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 370.39s
test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 2.36s

The test command was cargo test -p calternal-plugin-notes -- --test-threads=1 because independent test databases share the fixed fixture User ID and process-wide writer registry. The held-reconcile read test still uses its unchanged 100 ms budget and passed. Files clippy passed; full Files tests and the release server build remain in progress.

Round 4 desktop finding: `loadChunk` in the Calendar route waits for `Promise.all([loadRange(...), loadItems(...)])` before it fills `days`. The retained range alone does not contain an explicitly completed standalone rail. A warm Week/Day remount therefore still starts empty and awaits the chunk read. The view now retains a completed chunk with its filtered rail in the existing bounded, generation-aware Calendar cache. Initial grid state reads this snapshot synchronously. The publisher captures the request's zone, photo preference and cache generation before awaiting; a completed old request cannot refill a changed or signed-out User's cache. An items error does not replace a good completed snapshot. Three new tests cover complete rail retention, photo preference separation, auth reset and invalidation races. The data and Journal targeted tests pass (24 tests). Final web gates are running for this last client slice. Notes final gates passed: ``` Finished `dev` profile [unoptimized + debuginfo] target(s) in 14m 10s test result: ok. 132 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 370.39s test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 2.36s ``` The test command was `cargo test -p calternal-plugin-notes -- --test-threads=1` because independent test databases share the fixed fixture User ID and process-wide writer registry. The held-reconcile read test still uses its unchanged 100 ms budget and passed. Files clippy passed; full Files tests and the release server build remain in progress.
kayg commented 2026-10-01 10:31:37 +00:00
Author
Owner
Copy link

Round 4 finding: the full Calendar e2e reaches calendar.mjs:855 and fails the new log entry is in the Daily note (actual undefined). This is from this job: Composer batch ACKs queue the full projection, but the new committed Journal reader can still see the previous snapshot. The fix publishes the ready source snapshot before the batch ACK; Search, Calendar and DAV indexing remain queued. It does not wait for indexing on GET.

Regression: composer_batch_publishes_journal_before_queued_indexing holds the Notes guard after the batch ACK and requires the new Log through Journal GET within 100 ms. Output:

test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 132 filtered out; finished in 0.62s

The final Notes/Files/server gates are running. Files' first full test run had one unchanged five-minute storm timeout on the loaded host; I am running the crate serially, with the same assertions and timeout. The web gates pass with a 15-second runner budget: 142 files, 951 tests; no test expectations changed.

Decision: publish only the minimal ready Daily note source before the batch ACK. Keep the heavier queued projections. The benchmark now covers single-Log ACKs, a bounded 1,000-Log batch and 20 reads of that large day. It records browser-host load separately from perf-VM load.

Round 4 finding: the full Calendar e2e reaches calendar.mjs:855 and fails `the new log entry is in the Daily note` (actual undefined). This is from this job: Composer batch ACKs queue the full projection, but the new committed Journal reader can still see the previous snapshot. The fix publishes the ready source snapshot before the batch ACK; Search, Calendar and DAV indexing remain queued. It does not wait for indexing on GET. Regression: `composer_batch_publishes_journal_before_queued_indexing` holds the Notes guard after the batch ACK and requires the new Log through Journal GET within 100 ms. Output: ``` test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 132 filtered out; finished in 0.62s ``` The final Notes/Files/server gates are running. Files' first full test run had one unchanged five-minute storm timeout on the loaded host; I am running the crate serially, with the same assertions and timeout. The web gates pass with a 15-second runner budget: 142 files, 951 tests; no test expectations changed. Decision: publish only the minimal ready Daily note source before the batch ACK. Keep the heavier queued projections. The benchmark now covers single-Log ACKs, a bounded 1,000-Log batch and 20 reads of that large day. It records browser-host load separately from perf-VM load.
kayg commented 2026-10-01 10:57:33 +00:00
Author
Owner
Copy link

Round 4 client finding (local, not an HDD comparison): the latest current production UI restores cached Agenda and grid snapshots synchronously, enters the Tab route without awaiting preload, releases registered entry-route headers in the navigation update, and tracks label transitions through events instead of getAnimations. A restored Agenda also skips arrival motion for its initial rows; new rows and fresh views keep the shared motion.

Local Chromium 390 warm Photos → Calendar n DOM p50/p95 ms First p50/p95 ms Full p50/p95 ms CLS after DOM paint p95 Critical request-depth upper bound
Before restored-row motion change 11 360.9 / 1047.6 563.6 / 1350.0 608.4 / 1414.4 0 2
After restored-row motion change 11 290.2 / 614.9 422.4 / 955.8 472.0 / 1001.4 0 2

Host load and background work differ, so this is diagnostic evidence, not a controlled regression ratio. The 100 ms target remains unmet. The first trace has a 422.7 ms callback, a 92.5 ms layout of 3,924 objects and a 46.9 ms style pass. The later trace's longest task is 257.97 ms. Cached rows still cost DOM construction and layout; full row virtualization is not in this change. Web gates after the motion change: svelte-check found 0 errors and 0 warnings, Test Files 142 passed (142), Tests 951 passed (951).

The original Calendar tus fixture stop is diagnosed and filed as #627. It returns upload destination changed, with no If-Match sent. A bounded test reproduces 412 after re-indexing the same parent directory following an unrelated child write. The full flow failed before Round 4 too. The early parent-mtime exception did not fix that flow and was reverted; Files retains its full destination checks. This is separate from the Notes epoch triage in #427.

The rebuilt final embedded UI/server is compiling. The HDD warm/cold matrix is next, followed by the seven-Tab matrix as time permits.

Round 4 client finding (local, not an HDD comparison): the latest current production UI restores cached Agenda and grid snapshots synchronously, enters the Tab route without awaiting preload, releases registered entry-route headers in the navigation update, and tracks label transitions through events instead of getAnimations. A restored Agenda also skips arrival motion for its initial rows; new rows and fresh views keep the shared motion. | Local Chromium 390 warm Photos → Calendar | n | DOM p50/p95 ms | First p50/p95 ms | Full p50/p95 ms | CLS after DOM paint p95 | Critical request-depth upper bound | | --- | ---: | --- | --- | --- | ---: | ---: | | Before restored-row motion change | 11 | 360.9 / 1047.6 | 563.6 / 1350.0 | 608.4 / 1414.4 | 0 | 2 | | After restored-row motion change | 11 | 290.2 / 614.9 | 422.4 / 955.8 | 472.0 / 1001.4 | 0 | 2 | Host load and background work differ, so this is diagnostic evidence, not a controlled regression ratio. The 100 ms target remains unmet. The first trace has a 422.7 ms callback, a 92.5 ms layout of 3,924 objects and a 46.9 ms style pass. The later trace's longest task is 257.97 ms. Cached rows still cost DOM construction and layout; full row virtualization is not in this change. Web gates after the motion change: `svelte-check found 0 errors and 0 warnings`, `Test Files 142 passed (142)`, `Tests 951 passed (951)`. The original Calendar tus fixture stop is diagnosed and filed as #627. It returns `upload destination changed`, with no If-Match sent. A bounded test reproduces 412 after re-indexing the same parent directory following an unrelated child write. The full flow failed before Round 4 too. The early parent-mtime exception did not fix that flow and was reverted; Files retains its full destination checks. This is separate from the Notes epoch triage in #427. The rebuilt final embedded UI/server is compiling. The HDD warm/cold matrix is next, followed by the seven-Tab matrix as time permits.
kayg commented 2026-10-01 11:53:04 +00:00
Author
Owner
Copy link

The first qualified HDD matrix is still above the target. It uses build b126ffe335, before the latest row and cache-window changes. Do not count it as evidence for the final client.

Latest client change, 44cc2e98d: extend the existing Agenda deferral to rows. The first window uses viewport height and the existing 40 px minimum row, with two spare rows. One shared row observer has a 200 px lead; existing day and repair observation keeps a 600 px lead. Rows stay mounted after visibility, selection or keyboard use, including after an insertion shifts their position. Each mounted list item has its full data position and set size for assistive input. Keyboard and block links mount their target before moving focus. No text, style or per-row geometry measurement was added. This is deferred mounting, not unmounting/recycling row virtualization.

Latest cache change: retain six completed grid chunks. Week can rest on any date (DESIGN §39), so its 35-day rendered/prefetched window can span six week chunks. Four could evict the visible week by completion order. The regression retains the visible revision after all six complete and verifies bounded eviction after a seventh.

Final web output:

svelte-check found 0 errors and 0 warnings
 Test Files  142 passed (142)
      Tests  955 passed (955)

The focused cache tests also pass:

 Test Files  1 passed (1)
      Tests  20 passed (20)

The embedded final client is rebuilding. I will repeat the cold/warm engine/width matrix for it with eleven warm and three cold samples per profile, then measure the other Tabs as time allows. Decisions: conservative viewport arithmetic, one observer per visibility purpose, retain mounted heights, and a six-chunk grid LRU. These extend the existing components and cache writer.

The first qualified HDD matrix is still above the target. It uses build b126ffe335719471658c069e16252b3a140c3a50, before the latest row and cache-window changes. Do not count it as evidence for the final client. Latest client change, 44cc2e98d: extend the existing Agenda deferral to rows. The first window uses viewport height and the existing 40 px minimum row, with two spare rows. One shared row observer has a 200 px lead; existing day and repair observation keeps a 600 px lead. Rows stay mounted after visibility, selection or keyboard use, including after an insertion shifts their position. Each mounted list item has its full data position and set size for assistive input. Keyboard and block links mount their target before moving focus. No text, style or per-row geometry measurement was added. This is deferred mounting, not unmounting/recycling row virtualization. Latest cache change: retain six completed grid chunks. Week can rest on any date (DESIGN §39), so its 35-day rendered/prefetched window can span six week chunks. Four could evict the visible week by completion order. The regression retains the visible revision after all six complete and verifies bounded eviction after a seventh. Final web output: ``` svelte-check found 0 errors and 0 warnings Test Files 142 passed (142) Tests 955 passed (955) ``` The focused cache tests also pass: ``` Test Files 1 passed (1) Tests 20 passed (20) ``` The embedded final client is rebuilding. I will repeat the cold/warm engine/width matrix for it with eleven warm and three cold samples per profile, then measure the other Tabs as time allows. Decisions: conservative viewport arithmetic, one observer per visibility purpose, retain mounted heights, and a six-chunk grid LRU. These extend the existing components and cache writer.
kayg commented 2026-10-01 12:11:05 +00:00
Author
Owner
Copy link

Round 4 measurement update (head 66a14a005; final client source 814258d16):

The pre-row-deferral HDD matrix recorded all 88 requested switches (Chromium/WebKit, 390/1440, eleven warm and eleven cold samples each). Its optional Journal write profile then timed out on the first single-Log POST at 60 seconds, so the checkpoint JSON remains in_progress. This is an incomplete write measurement, not a successful latency result. Journal reads all returned 200. The final row-deferral/cache-window client matrix is running with eleven warm and three actual cache-drop/server-restart cold samples per engine/width.

The first matrix's warm Calendar DOM p95 values were 4505.3 ms (Chromium390), 1604 ms (Chromium1440), 5433 ms (WebKit390), and 9118 ms (WebKit1440). Browser-host load ranged 5.68–34.32; no outliers are discarded, and this is not a controlled comparison to Round 3. These values do not meet the 100 ms target.

The serial Journal read took 42.884 ms end to end, with 1.519 ms plugin Server-Timing. The 20-read burst had p50 224.664 ms/p95 237.130 ms, 150 ms process CPU, 394055680 bytes RSS and no physical reads. This separates the remaining HTTP/session/pool cost from Notes reconciliation. The held-writer regression and the real-server reconciliation probe still pass.

A failed Playwright request prints credential headers unless they are redacted. The bench now redacts those lines as well as setup URLs, with an ANSI-header regression test (1 pass, 0 fail). Diagnostic logs were checked before attachment.

Round 4 measurement update (head 66a14a005; final client source 814258d16): The pre-row-deferral HDD matrix recorded all 88 requested switches (Chromium/WebKit, 390/1440, eleven warm and eleven cold samples each). Its optional Journal write profile then timed out on the first single-Log POST at 60 seconds, so the checkpoint JSON remains `in_progress`. This is an incomplete write measurement, not a successful latency result. Journal reads all returned 200. The final row-deferral/cache-window client matrix is running with eleven warm and three actual cache-drop/server-restart cold samples per engine/width. The first matrix's warm Calendar DOM p95 values were 4505.3 ms (Chromium390), 1604 ms (Chromium1440), 5433 ms (WebKit390), and 9118 ms (WebKit1440). Browser-host load ranged 5.68–34.32; no outliers are discarded, and this is not a controlled comparison to Round 3. These values do not meet the 100 ms target. The serial Journal read took 42.884 ms end to end, with 1.519 ms plugin Server-Timing. The 20-read burst had p50 224.664 ms/p95 237.130 ms, 150 ms process CPU, 394055680 bytes RSS and no physical reads. This separates the remaining HTTP/session/pool cost from Notes reconciliation. The held-writer regression and the real-server reconciliation probe still pass. A failed Playwright request prints credential headers unless they are redacted. The bench now redacts those lines as well as setup URLs, with an ANSI-header regression test (1 pass, 0 fail). Diagnostic logs were checked before attachment.
kayg commented 2026-10-01 12:34:45 +00:00
Author
Owner
Copy link

Final client HDD matrix recorded at 71597bfb6 (production client source 814258d160). All 56 samples completed. The 100 ms warm target remains unmet.

Final Round 4 client HDD matrix

The final production client source is 814258d16. The result is
runs/tab-switch-2026-10-01-549-round4-final-hdd.json.
It has 56 completed samples: eleven warm and three cold per browser and width.
Cold samples stop the server, drop Linux page caches and start the same binary.
Each measured run holds /root/perf.lock. The browser runs on the shared build
host. Light mode and the smoke Home are used; no outlier is removed.

HDD qualification again gives QD1 125.008 IOPS, p50 8.028 ms, p99 8.159 ms.
QD16 gives 200.906 IOPS, p50 100.139 ms, p99 104.333 ms. The VM load during
qualification is 1.17/1.89/2.19 and 1.40/1.91/2.19.
Browser-host one-minute load during switches ranges from 8.52 to 21.72.

Engine Width State n / DOM n DOM p50/p95 ms First p50/p95 ms Full p50/p95 ms CLS after DOM p95 Request-depth upper bound Long task p95 ms
chromium 390 warm 11 / 11 127.1/229.5 227.9/357.1 262.2/486.8 0 3 130
chromium 390 cold 3 / 3 458.7/568.5 199.3/354.6 241.1/687 0 3 106
chromium 1440 warm 11 / 11 642.9/1729.8 1530.5/2313.2 1833.3/3995.3 0 6 1074
chromium 1440 cold 3 / 3 780.7/1121.7 1004.8/1723.1 2165.6/3057.7 0 9 482
webkit 390 warm 11 / 11 541/851 675/1092 724/1202 unsupported 2 unsupported
webkit 390 cold 3 / 3 646/684 401/484 766/888 unsupported 5 unsupported
webkit 1440 warm 11 / 11 1565/2159 2111/2983 2310/3388 unsupported 10 unsupported
webkit 1440 cold 3 / 3 1779/2784 1989/3916 2850/4493 unsupported 13 unsupported

Three cold samples are a bounded confirmation, not a strong percentile
estimate. The historic First marker can precede the real Calendar DOM marker.
Use the DOM marker for real Calendar content; Full also waits for visible
loading UI, fonts and image decoding. WebKit CLS and long tasks are unavailable.
Standard CLS excludes recent-input shifts. Raw shift sources are in the JSON.

Engine Width State Server CPU p50/p95 ms Server RSS p50/p95 bytes Physical read bytes p50 VM load p50/p95
chromium 390 warm 170/460 185389056/540561408 21217280 3.21/3.56
chromium 390 cold 640/660 166322176/166776832 46002176 3.22/3.56
chromium 1440 warm 370/970 421408768/439566336 1908736 2.83/3.01
chromium 1440 cold 1000/1020 170860544/172445696 67190784 2.64/2.78
webkit 390 warm 260/560 393334784/448540672 1286144 2.58/2.72
webkit 390 cold 730/760 167211008/167301120 45670400 2.47/2.78
webkit 1440 warm 380/1510 419983360/425541632 2035712 2.26/2.86
webkit 1440 cold 1100/1160 170090496/171786240 81367040 2.25/2.38

Final production header e2e: calendar header e2e: all geometry checks passed. There are 36 screenshots across 390/820/1440 px, light/dark, Agenda/Day/Week/Log editing. Claude reviews visual quality.

Final-server adversarial output:

Journal reconcile burst: n=20 p50=42.3 ms p95=43.0 ms
calendar Event tag probe: Unicode/bidi, 65536-byte category, 10 malformed inputs, 24 parallel tag/range/search reads, and Calendar/Notes/Files Server-Timing passed

The remaining all-Tab run measures first visits for every enabled Tab at all six width/theme profiles, plus one warm return to Calendar from each other Tab. These single-sample diagnostics do not replace the eleven-sample matrix.

Final client HDD matrix recorded at 71597bfb6 (production client source 814258d16089ac289f458f8a37edf907857b9d9a). All 56 samples completed. The 100 ms warm target remains unmet. ### Final Round 4 client HDD matrix The final production client source is `814258d16`. The result is [`runs/tab-switch-2026-10-01-549-round4-final-hdd.json`](runs/tab-switch-2026-10-01-549-round4-final-hdd.json). It has 56 completed samples: eleven warm and three cold per browser and width. Cold samples stop the server, drop Linux page caches and start the same binary. Each measured run holds `/root/perf.lock`. The browser runs on the shared build host. Light mode and the smoke Home are used; no outlier is removed. HDD qualification again gives QD1 125.008 IOPS, p50 8.028 ms, p99 8.159 ms. QD16 gives 200.906 IOPS, p50 100.139 ms, p99 104.333 ms. The VM load during qualification is 1.17/1.89/2.19 and 1.40/1.91/2.19. Browser-host one-minute load during switches ranges from 8.52 to 21.72. | Engine | Width | State | n / DOM n | DOM p50/p95 ms | First p50/p95 ms | Full p50/p95 ms | CLS after DOM p95 | Request-depth upper bound | Long task p95 ms | | --- | ---: | --- | --- | --- | --- | --- | ---: | ---: | ---: | | chromium | 390 | warm | 11 / 11 | 127.1/229.5 | 227.9/357.1 | 262.2/486.8 | 0 | 3 | 130 | | chromium | 390 | cold | 3 / 3 | 458.7/568.5 | 199.3/354.6 | 241.1/687 | 0 | 3 | 106 | | chromium | 1440 | warm | 11 / 11 | 642.9/1729.8 | 1530.5/2313.2 | 1833.3/3995.3 | 0 | 6 | 1074 | | chromium | 1440 | cold | 3 / 3 | 780.7/1121.7 | 1004.8/1723.1 | 2165.6/3057.7 | 0 | 9 | 482 | | webkit | 390 | warm | 11 / 11 | 541/851 | 675/1092 | 724/1202 | unsupported | 2 | unsupported | | webkit | 390 | cold | 3 / 3 | 646/684 | 401/484 | 766/888 | unsupported | 5 | unsupported | | webkit | 1440 | warm | 11 / 11 | 1565/2159 | 2111/2983 | 2310/3388 | unsupported | 10 | unsupported | | webkit | 1440 | cold | 3 / 3 | 1779/2784 | 1989/3916 | 2850/4493 | unsupported | 13 | unsupported | Three cold samples are a bounded confirmation, not a strong percentile estimate. The historic First marker can precede the real Calendar DOM marker. Use the DOM marker for real Calendar content; Full also waits for visible loading UI, fonts and image decoding. WebKit CLS and long tasks are unavailable. Standard CLS excludes recent-input shifts. Raw shift sources are in the JSON. | Engine | Width | State | Server CPU p50/p95 ms | Server RSS p50/p95 bytes | Physical read bytes p50 | VM load p50/p95 | | --- | ---: | --- | --- | --- | ---: | --- | | chromium | 390 | warm | 170/460 | 185389056/540561408 | 21217280 | 3.21/3.56 | | chromium | 390 | cold | 640/660 | 166322176/166776832 | 46002176 | 3.22/3.56 | | chromium | 1440 | warm | 370/970 | 421408768/439566336 | 1908736 | 2.83/3.01 | | chromium | 1440 | cold | 1000/1020 | 170860544/172445696 | 67190784 | 2.64/2.78 | | webkit | 390 | warm | 260/560 | 393334784/448540672 | 1286144 | 2.58/2.72 | | webkit | 390 | cold | 730/760 | 167211008/167301120 | 45670400 | 2.47/2.78 | | webkit | 1440 | warm | 380/1510 | 419983360/425541632 | 2035712 | 2.26/2.86 | | webkit | 1440 | cold | 1100/1160 | 170090496/171786240 | 81367040 | 2.25/2.38 | Final production header e2e: `calendar header e2e: all geometry checks passed`. There are 36 screenshots across 390/820/1440 px, light/dark, Agenda/Day/Week/Log editing. Claude reviews visual quality. Final-server adversarial output: ``` Journal reconcile burst: n=20 p50=42.3 ms p95=43.0 ms calendar Event tag probe: Unicode/bidi, 65536-byte category, 10 malformed inputs, 24 parallel tag/range/search reads, and Calendar/Notes/Files Server-Timing passed ``` The remaining all-Tab run measures first visits for every enabled Tab at all six width/theme profiles, plus one warm return to Calendar from each other Tab. These single-sample diagnostics do not replace the eleven-sample matrix.
kayg commented 2026-10-01 12:44:17 +00:00
Author
Owner
Copy link

Paint-boundary finding (738405d2f): pending Tab selection changes before route commit. The old benchmark checked only a non-empty route root, so a cold Calendar Full result could still measure Photos. Calendar DOM marks check Calendar content independently; their warm p95 values remain useful. The old First/Full and request-depth figures are now marked invalid in the report, and their CLS window can be shorter than one second after target paint.

The harness now waits for the target pathname and Calendar root before First/Full. Regression output:

 2 pass
 0 fail
 15 expect() calls
Ran 2 tests across 1 file. [121.00ms]

I stopped the all-Tab run during setup (no samples) to prioritize a corrected HDD warm/cold smoke: one sample per condition, Chromium/WebKit and 390/1440. Eleven warm DOM samples per profile remain in the earlier final-client matrix; the corrected smoke verifies the target Full boundary, rather than fabricating a p95 from one sample. Final screenshots and the real-server adversarial probe already pass. The 100 ms warm target remains unmet.

Paint-boundary finding (738405d2f): pending Tab selection changes before route commit. The old benchmark checked only a non-empty route root, so a cold Calendar Full result could still measure Photos. Calendar DOM marks check Calendar content independently; their warm p95 values remain useful. The old First/Full and request-depth figures are now marked invalid in the report, and their CLS window can be shorter than one second after target paint. The harness now waits for the target pathname and Calendar root before First/Full. Regression output: ``` 2 pass 0 fail 15 expect() calls Ran 2 tests across 1 file. [121.00ms] ``` I stopped the all-Tab run during setup (no samples) to prioritize a corrected HDD warm/cold smoke: one sample per condition, Chromium/WebKit and 390/1440. Eleven warm DOM samples per profile remain in the earlier final-client matrix; the corrected smoke verifies the target Full boundary, rather than fabricating a p95 from one sample. Final screenshots and the real-server adversarial probe already pass. The 100 ms warm target remains unmet.
kayg commented 2026-10-01 13:02:52 +00:00
Author
Owner
Copy link

Round 4 final report. Head: c19f56116032da92c1f432c2b60358e5d94657e3 on job/tabswitch-549. Initial base: 2c10cc92f7ce1bf001cc4dcdad8a3c549c2e8931. One authorized merge of origin/dev: 307d8221a1f34d59944c67c355a67baa46adb8fc (origin/dev parent 4dc9ca394). No push or deploy. Working tree is clean.

Result: the warm 100 ms p95 goal remains unmet. Final Chromium390 Calendar DOM p50/p95 is 127.1/229.5 ms across eleven samples. The final trace's longest post-pointerdown task is 65.58 ms; the eleven-sample longest-task p95 is 130 ms. Round 3 optimizations remain. The prior qualified DOM p95 was 768 ms, but host loads differ, so no controlled ratio is claimed.

Built

  • Notes migration 24 stores complete Journal source snapshots. Normal day GET uses WAL and serves the previous committed source while reconciliation holds the Notes guard and SQLite writer. A regression requires the read within 100 ms and checks the new source after commit.
  • Composer batch publishes ready Journal source before ACK while full Search/Calendar/DAV indexing stays queued. A regression requires the acknowledged Log while the Notes guard remains held. This fixes the earlier full Calendar stop at calendar.mjs:855.
  • Calendar mounts cached Agenda and complete Day/Week snapshots before any await, retains stale snapshots for guarded refresh, and keeps six completed weekly chunks including the saved-item rail and photo-time key. Auth changes clear them.
  • Tab navigation starts before preload completes; entry redirects release the previous header in the same update. Label morph uses transition events instead of getAnimations. Restored rows skip initial arrival motion.
  • Dense Agenda mounts a bounded initial row window and uses one shared observer for later rows. Keyboard and block reveal mount their destination before focus; mounted rows remain.
  • Log PATCH/DELETE and repair POST obtain current revisions immediately before If-Match. Cached Calendar snapshots supply no write precondition.
  • Extended the benchmark with server-phase/burst/write profiles, browser-host load, post-paint CLS, and a corrected target-route wait. Added diagnostic redaction and two regressions. Extended the real-server defensive Journal probe.

412 diagnosis

The final full Calendar flow still reaches calendar.mjs:1117 and receives 412 upload destination changed instead of 201. It sends no If-Match. Parent reindex after a sibling write can change the Files directory item ID. #627 has the reproducer. This predates Round 4 and is separate from #427 Notes epoch triage. The attempted parent-mtime exception was reverted because it did not fix the real flow; final Files behavior is unchanged. No existing e2e expectation or fixture was weakened. Final full-flow proof.

Files

crates/plugins/notes/src/lib.rs
crates/plugins/notes/src/store.rs
crates/plugins/notes/migrations/0024_journal_day_snapshots.sql
apps/web/src/lib/calendar/journal.ts
apps/web/src/lib/calendar/journal.test.ts
apps/web/src/lib/calendar/data.ts
apps/web/src/lib/calendar/data.test.ts
apps/web/src/lib/calendar/agenda.svelte.test.ts
apps/web/src/lib/tray.svelte.test.ts
apps/web/src/routes/+layout.svelte
apps/web/src/routes/calendar/[view]/[date]/+page.svelte
packages/ui/src/components/SegmentedControl.svelte
packages/ui/src/components/calendar/AgendaList.svelte
bench/tab-switch.mjs
bench/tab-switch.test.mjs
bench/run.sh
tests/adversarial/calendar_event_tags.mjs
docs/perf/2026-10-01-tabswitch-549.md
docs/perf/runs/tab-switch-2026-10-01-549-round4-final-hdd.json
docs/perf/runs/tab-switch-2026-10-01-549-round4-corrected-smoke.json

Measurements

The final production source is 814258d16089ac289f458f8a37edf907857b9d9a. The 56-sample HDD matrix has eleven warm and three actual server-restart/Linux-cache-drop cold samples per engine/width. It uses light mode and the smoke Home (366 Daily notes, 10980 Logs, 30 daily recurring Events, 100 Photos). Every measured phase holds /root/perf.lock; load is recorded inside it. QD1: 125.008 IOPS, p50 8.028 ms, p99 8.159 ms. QD16: 200.906 IOPS, p50 100.139 ms, p99 104.333 ms. Browser-host switch load ranges 8.52–21.72. No outlier is removed.

The benchmark review found that pending Tab selection could pass the old non-empty-root check on Photos. The older First/Full and request-depth counters are invalid target measurements; their CLS window can be shorter than one second after target paint. The Calendar DOM marks check Calendar content independently. The saved report qualifies these old counters. The corrected smoke below requires target pathname and Calendar root before First/Full.

Engine Width Warm n Calendar DOM p50/p95 ms Longest task p95 ms
chromium 390 11 127.1/229.5 130
chromium 1440 11 642.9/1729.8 1074
webkit 390 11 541/851 unsupported
webkit 1440 11 1565/2159 unsupported

Corrected HDD smoke: eight completed samples, one per row. These are diagnostic values, not p95 estimates. Cold includes Linux cache drop plus server restart. CLS observes one second after target Full. Request depth is a chronological non-overlap upper bound, not a causal graph.

Engine Width State n DOM ms First ms Full ms CLS after DOM Request-depth upper bound CPU ms RSS bytes
chromium 390 cold 1 339.3 391.9 419.4 0 4 680 171356160
chromium 390 warm 1 130.4 202.3 249.8 0 3 1020 181235712
chromium 1440 cold 1 780.7 1334.8 2459.8 0 8 840 168796160
chromium 1440 warm 1 438.4 1100.8 1245.2 0 7 780 183521280
webkit 390 cold 1 593 837 879 unsupported 4 630 162594816
webkit 390 warm 1 370 589 787 unsupported 3 780 177274880
webkit 1440 cold 1 810 1406 1879 unsupported 13 960 173211648
webkit 1440 warm 1 1435 2107 2226 unsupported 10 390 188473344

Corrected serial Journal read: 41.282 ms end to end, 0.837 ms plugin phases. Twenty-read burst: p50 126.520 ms/p95 156.732 ms, CPU 280 ms, RSS 391319552 bytes, no physical reads. Calendar range: 74.762 ms end to end, 12.207 ms phases, 1.953 ms database. Snapshot reads have no Notes lock phase. Earlier loaded probes and process CPU/RSS tables remain in the dated report. baseline.json has no matching snapshot/batch/Tab profile; its Calendar range p95 is 4.0 ms and Notes Daily p95 4.8 ms, with different workloads. No baseline is promoted.

The warm trace starts its large Svelte queue callback 62.09 ms after pointerdown; it takes 57.66 ms. Layout starts at 133.34 ms and takes 7.50 ms over 1248 objects (earlier trace: 3924). Navigation and component construction remain on the path to paint.

Gates (verbatim result excerpts; full logs attached)

cargo fmt --check: exit 0, no output. cargo clippy -p <crate> --all-targets -- -D warnings and cargo test -p <crate> passed for calternal-plugin-notes, restored calternal-plugin-files, and calternal-server. Notes and Files tests were serial because fixed process-global fixtures contend. Web commands: bun run --cwd apps/web check, bun run --cwd apps/web test --testTimeout 15000, production build. The CLI budget handles shared-host SLOW; assertions are unchanged. Benchmark command: bun test bench/tab-switch.test.mjs.

round4-final-fmt.log
(no output; exit 0)

round4-final-notes-clippy-current.log
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 3m 34s

round4-final-notes-test-current.log
    Finished `test` profile [unoptimized + debuginfo] target(s) in 26.33s
test result: ok. 133 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 209.86s
test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 1.03s
test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s

round4-restored-files-clippy.log
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 23.75s

round4-restored-files-test.log
    Finished `test` profile [unoptimized + debuginfo] target(s) in 1m 11s
test result: ok. 146 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 308.49s
test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s

round4-restored-server-clippy.log
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 1m 15s

round4-restored-server-test.log
    Finished `test` profile [unoptimized + debuginfo] target(s) in 7m 18s
test result: ok. 93 passed; 0 failed; 3 ignored; 0 measured; 0 filtered out; finished in 22.77s

round4-grid-bound-web-check.log
svelte-check found 0 errors and 0 warnings

round4-grid-bound-web-test.log
 Test Files  142 passed (142)
      Tests  955 passed (955)
   Duration  149.24s (transform 56%, environment 16%, import 14%, tests 10%, setup 3%)

round4-bench-target-test.log
 2 pass
 0 fail
 15 expect() calls
Ran 2 tests across 1 file. [121.00ms]

Final production header output:

calendar header e2e: all geometry checks passed

Final real-server adversarial output:

Journal reconcile burst: n=20 p50=42.3 ms p95=43.0 ms
calendar Event tag probe: Unicode/bidi, 65536-byte category, 10 malformed inputs, 24 parallel tag/range/search reads, and Calendar/Notes/Files Server-Timing passed

Final review bundle: 36 production Calendar screenshots for Agenda/Day/Week/Log editing at 390/820/1440 px in both schemes, including 3x header crops; dense HDD screenshots and trace; gates; raw measurements. Claude reviews visual quality. The bundle's earlier full-calendar-e2e log records the now-fixed line855 failure; the separate final proof above records the remaining line1117 412.

Known gaps

  • Warm 100 ms p95 is unmet; desktop and WebKit remain slower. Corrected cold is one sample per profile, so cold p95 is not verified.
  • Full Calendar e2e is blocked by existing Files tus #627. Notes epoch #427 is not changed.
  • Legacy empty-snapshot ID repair is idle-only but can wait on another User's SQLite writer. The strict 100 ms guarantee is proved for normal complete day snapshots, not that compatibility path.
  • Largest three-year/50k Photos/100k Files Home and all seven Tabs remain unmeasured. All-Tab setup was stopped to prioritize the corrected cold boundary. Photos first visits were measured in the Calendar matrix.
  • WebKit CLS/long-task observers are unavailable; true causal request depth is not computed. Row recycling is not implemented.
  • Optional Journal write profile timed out on its first single-Log POST at 60 seconds after the earlier 88-switch matrix. No write/1000-Log/large-day burst percentile is claimed.

Decisions not specified by DESIGN

Use complete Journal source snapshots and minimal ready batch publication; preserve idle-only legacy repair for compatibility. Keep six completed weekly chunks for the 35-day rendered/prefetched window. Mount max(12, ceil(viewport height/40)+2) Agenda rows with a shared 200 px observer lead and retain mounted rows. Use eleven warm/three cold samples, then a corrected one-sample boundary confirmation within the time limit. Observe CLS for one second after Full; report request-depth limits explicitly. Keep the shared baseline unchanged.

All touched module and non-obvious function comments were re-read. Build cleanup completed:

     Removed 25182 files, 10.9GiB total

Web build output and this worktree's temporary test/build files were removed. Review artifacts remain ignored. Issues stay open.

Round 4 final report. Head: `c19f56116032da92c1f432c2b60358e5d94657e3` on `job/tabswitch-549`. Initial base: `2c10cc92f7ce1bf001cc4dcdad8a3c549c2e8931`. One authorized merge of `origin/dev`: `307d8221a1f34d59944c67c355a67baa46adb8fc` (origin/dev parent `4dc9ca394`). No push or deploy. Working tree is clean. **Result: the warm 100 ms p95 goal remains unmet.** Final Chromium390 Calendar DOM p50/p95 is 127.1/229.5 ms across eleven samples. The final trace's longest post-pointerdown task is 65.58 ms; the eleven-sample longest-task p95 is 130 ms. Round 3 optimizations remain. The prior qualified DOM p95 was 768 ms, but host loads differ, so no controlled ratio is claimed. **Built** - Notes migration 24 stores complete Journal source snapshots. Normal day GET uses WAL and serves the previous committed source while reconciliation holds the Notes guard and SQLite writer. A regression requires the read within 100 ms and checks the new source after commit. - Composer batch publishes ready Journal source before ACK while full Search/Calendar/DAV indexing stays queued. A regression requires the acknowledged Log while the Notes guard remains held. This fixes the earlier full Calendar stop at calendar.mjs:855. - Calendar mounts cached Agenda and complete Day/Week snapshots before any await, retains stale snapshots for guarded refresh, and keeps six completed weekly chunks including the saved-item rail and photo-time key. Auth changes clear them. - Tab navigation starts before preload completes; entry redirects release the previous header in the same update. Label morph uses transition events instead of getAnimations. Restored rows skip initial arrival motion. - Dense Agenda mounts a bounded initial row window and uses one shared observer for later rows. Keyboard and block reveal mount their destination before focus; mounted rows remain. - Log PATCH/DELETE and repair POST obtain current revisions immediately before If-Match. Cached Calendar snapshots supply no write precondition. - Extended the benchmark with server-phase/burst/write profiles, browser-host load, post-paint CLS, and a corrected target-route wait. Added diagnostic redaction and two regressions. Extended the real-server defensive Journal probe. **412 diagnosis** The final full Calendar flow still reaches calendar.mjs:1117 and receives 412 `upload destination changed` instead of 201. It sends no If-Match. Parent reindex after a sibling write can change the Files directory item ID. [#627](https://git.kayg.org/kayg/calternal/issues/627) has the reproducer. This predates Round 4 and is separate from #427 Notes epoch triage. The attempted parent-mtime exception was reverted because it did not fix the real flow; final Files behavior is unchanged. No existing e2e expectation or fixture was weakened. [Final full-flow proof](https://git.kayg.org/attachments/55ddba03-e93f-4782-965e-58cb14a46dca). **Files** ``` crates/plugins/notes/src/lib.rs crates/plugins/notes/src/store.rs crates/plugins/notes/migrations/0024_journal_day_snapshots.sql apps/web/src/lib/calendar/journal.ts apps/web/src/lib/calendar/journal.test.ts apps/web/src/lib/calendar/data.ts apps/web/src/lib/calendar/data.test.ts apps/web/src/lib/calendar/agenda.svelte.test.ts apps/web/src/lib/tray.svelte.test.ts apps/web/src/routes/+layout.svelte apps/web/src/routes/calendar/[view]/[date]/+page.svelte packages/ui/src/components/SegmentedControl.svelte packages/ui/src/components/calendar/AgendaList.svelte bench/tab-switch.mjs bench/tab-switch.test.mjs bench/run.sh tests/adversarial/calendar_event_tags.mjs docs/perf/2026-10-01-tabswitch-549.md docs/perf/runs/tab-switch-2026-10-01-549-round4-final-hdd.json docs/perf/runs/tab-switch-2026-10-01-549-round4-corrected-smoke.json ``` **Measurements** The final production source is `814258d16089ac289f458f8a37edf907857b9d9a`. The 56-sample HDD matrix has eleven warm and three actual server-restart/Linux-cache-drop cold samples per engine/width. It uses light mode and the smoke Home (366 Daily notes, 10980 Logs, 30 daily recurring Events, 100 Photos). Every measured phase holds `/root/perf.lock`; load is recorded inside it. QD1: 125.008 IOPS, p50 8.028 ms, p99 8.159 ms. QD16: 200.906 IOPS, p50 100.139 ms, p99 104.333 ms. Browser-host switch load ranges 8.52–21.72. No outlier is removed. The benchmark review found that pending Tab selection could pass the old non-empty-root check on Photos. The older First/Full and request-depth counters are invalid target measurements; their CLS window can be shorter than one second after target paint. The Calendar DOM marks check Calendar content independently. The saved report qualifies these old counters. The corrected smoke below requires target pathname and Calendar root before First/Full. | Engine | Width | Warm n | Calendar DOM p50/p95 ms | Longest task p95 ms | | --- | ---: | ---: | --- | ---: | | chromium | 390 | 11 | 127.1/229.5 | 130 | | chromium | 1440 | 11 | 642.9/1729.8 | 1074 | | webkit | 390 | 11 | 541/851 | unsupported | | webkit | 1440 | 11 | 1565/2159 | unsupported | Corrected HDD smoke: eight completed samples, one per row. These are diagnostic values, not p95 estimates. Cold includes Linux cache drop plus server restart. CLS observes one second after target Full. Request depth is a chronological non-overlap upper bound, not a causal graph. | Engine | Width | State | n | DOM ms | First ms | Full ms | CLS after DOM | Request-depth upper bound | CPU ms | RSS bytes | | --- | ---: | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | | chromium | 390 | cold | 1 | 339.3 | 391.9 | 419.4 | 0 | 4 | 680 | 171356160 | | chromium | 390 | warm | 1 | 130.4 | 202.3 | 249.8 | 0 | 3 | 1020 | 181235712 | | chromium | 1440 | cold | 1 | 780.7 | 1334.8 | 2459.8 | 0 | 8 | 840 | 168796160 | | chromium | 1440 | warm | 1 | 438.4 | 1100.8 | 1245.2 | 0 | 7 | 780 | 183521280 | | webkit | 390 | cold | 1 | 593 | 837 | 879 | unsupported | 4 | 630 | 162594816 | | webkit | 390 | warm | 1 | 370 | 589 | 787 | unsupported | 3 | 780 | 177274880 | | webkit | 1440 | cold | 1 | 810 | 1406 | 1879 | unsupported | 13 | 960 | 173211648 | | webkit | 1440 | warm | 1 | 1435 | 2107 | 2226 | unsupported | 10 | 390 | 188473344 | Corrected serial Journal read: 41.282 ms end to end, 0.837 ms plugin phases. Twenty-read burst: p50 126.520 ms/p95 156.732 ms, CPU 280 ms, RSS 391319552 bytes, no physical reads. Calendar range: 74.762 ms end to end, 12.207 ms phases, 1.953 ms database. Snapshot reads have no Notes lock phase. Earlier loaded probes and process CPU/RSS tables remain in the dated report. `baseline.json` has no matching snapshot/batch/Tab profile; its Calendar range p95 is 4.0 ms and Notes Daily p95 4.8 ms, with different workloads. No baseline is promoted. The warm trace starts its large Svelte queue callback 62.09 ms after pointerdown; it takes 57.66 ms. Layout starts at 133.34 ms and takes 7.50 ms over 1248 objects (earlier trace: 3924). Navigation and component construction remain on the path to paint. **Gates (verbatim result excerpts; full logs attached)** `cargo fmt --check`: exit 0, no output. `cargo clippy -p <crate> --all-targets -- -D warnings` and `cargo test -p <crate>` passed for calternal-plugin-notes, restored calternal-plugin-files, and calternal-server. Notes and Files tests were serial because fixed process-global fixtures contend. Web commands: `bun run --cwd apps/web check`, `bun run --cwd apps/web test --testTimeout 15000`, production build. The CLI budget handles shared-host SLOW; assertions are unchanged. Benchmark command: `bun test bench/tab-switch.test.mjs`. ``` round4-final-fmt.log (no output; exit 0) round4-final-notes-clippy-current.log Finished `dev` profile [unoptimized + debuginfo] target(s) in 3m 34s round4-final-notes-test-current.log Finished `test` profile [unoptimized + debuginfo] target(s) in 26.33s test result: ok. 133 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 209.86s test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 1.03s test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s round4-restored-files-clippy.log Finished `dev` profile [unoptimized + debuginfo] target(s) in 23.75s round4-restored-files-test.log Finished `test` profile [unoptimized + debuginfo] target(s) in 1m 11s test result: ok. 146 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 308.49s test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s round4-restored-server-clippy.log Finished `dev` profile [unoptimized + debuginfo] target(s) in 1m 15s round4-restored-server-test.log Finished `test` profile [unoptimized + debuginfo] target(s) in 7m 18s test result: ok. 93 passed; 0 failed; 3 ignored; 0 measured; 0 filtered out; finished in 22.77s round4-grid-bound-web-check.log svelte-check found 0 errors and 0 warnings round4-grid-bound-web-test.log Test Files 142 passed (142) Tests 955 passed (955) Duration 149.24s (transform 56%, environment 16%, import 14%, tests 10%, setup 3%) round4-bench-target-test.log 2 pass 0 fail 15 expect() calls Ran 2 tests across 1 file. [121.00ms] ``` Final production header output: ``` calendar header e2e: all geometry checks passed ``` Final real-server adversarial output: ``` Journal reconcile burst: n=20 p50=42.3 ms p95=43.0 ms calendar Event tag probe: Unicode/bidi, 65536-byte category, 10 malformed inputs, 24 parallel tag/range/search reads, and Calendar/Notes/Files Server-Timing passed ``` [Final review bundle](https://git.kayg.org/attachments/a48d46be-52e9-4bc6-8a6c-151e6b7d045c): 36 production Calendar screenshots for Agenda/Day/Week/Log editing at 390/820/1440 px in both schemes, including 3x header crops; dense HDD screenshots and trace; gates; raw measurements. Claude reviews visual quality. The bundle's earlier full-calendar-e2e log records the now-fixed line855 failure; the separate final proof above records the remaining line1117 412. **Known gaps** - Warm 100 ms p95 is unmet; desktop and WebKit remain slower. Corrected cold is one sample per profile, so cold p95 is not verified. - Full Calendar e2e is blocked by existing Files tus #627. Notes epoch #427 is not changed. - Legacy empty-snapshot ID repair is idle-only but can wait on another User's SQLite writer. The strict 100 ms guarantee is proved for normal complete day snapshots, not that compatibility path. - Largest three-year/50k Photos/100k Files Home and all seven Tabs remain unmeasured. All-Tab setup was stopped to prioritize the corrected cold boundary. Photos first visits were measured in the Calendar matrix. - WebKit CLS/long-task observers are unavailable; true causal request depth is not computed. Row recycling is not implemented. - Optional Journal write profile timed out on its first single-Log POST at 60 seconds after the earlier 88-switch matrix. No write/1000-Log/large-day burst percentile is claimed. **Decisions not specified by DESIGN** Use complete Journal source snapshots and minimal ready batch publication; preserve idle-only legacy repair for compatibility. Keep six completed weekly chunks for the 35-day rendered/prefetched window. Mount max(12, ceil(viewport height/40)+2) Agenda rows with a shared 200 px observer lead and retain mounted rows. Use eleven warm/three cold samples, then a corrected one-sample boundary confirmation within the time limit. Observe CLS for one second after Full; report request-depth limits explicitly. Keep the shared baseline unchanged. All touched module and non-obvious function comments were re-read. Build cleanup completed: ``` Removed 25182 files, 10.9GiB total ``` Web build output and this worktree's temporary test/build files were removed. Review artifacts remain ignored. Issues stay open.
kayg commented 2026-10-02 06:35:28 +00:00
Author
Owner
Copy link

Starting measurement-only rerun for #549 on branch job/perf-rerun.

Base and comparison baseline: origin/dev at c4a61e8cf090170f35b1bed3350d9de20c83ecd5 (merge round 6). I will use the shared release artifacts, qualify HDD emulation with fio, hold /root/perf.lock for every measured run, and record uptime/load with interleaved A/B samples. Scope is the six requested profiles (#632, #573, #476, #651, #549, #641); no product code changes.

Starting measurement-only rerun for #549 on branch `job/perf-rerun`. Base and comparison baseline: `origin/dev` at `c4a61e8cf090170f35b1bed3350d9de20c83ecd5` (merge round 6). I will use the shared release artifacts, qualify HDD emulation with fio, hold `/root/perf.lock` for every measured run, and record uptime/load with interleaved A/B samples. Scope is the six requested profiles (#632, #573, #476, #651, #549, #641); no product code changes.
kayg commented 2026-10-02 06:39:40 +00:00
Author
Owner
Copy link

HDD qualification completed under flock -w 14400 /root/perf.lock on the fresh VM boot. bench/hdd-emu.sh up mounted /srv/hdd-emu from /dev/mapper/calternal-hdd-emu (direct-I/O loop, 8 ms read/write delay). Command: fio --name=hdd-qualification --filename=/srv/hdd-emu/qualification.bin --size=128m --rw=randread --bs=4k --ioengine=libaio --iodepth=1 --numjobs=1 --direct=1 --time_based=1 --runtime=30 --ramp_time=3 --group_reporting=1 --output-format=json. Result: 117.01 IOPS, p50 7.96 ms, p99 16.58 ms. Uptime at start: 3:35; load average: 0.04, 0.02, 0.07. The IOPS and median latency match the expected HDD profile.

HDD qualification completed under `flock -w 14400 /root/perf.lock` on the fresh VM boot. `bench/hdd-emu.sh up` mounted `/srv/hdd-emu` from `/dev/mapper/calternal-hdd-emu` (direct-I/O loop, 8 ms read/write delay). Command: `fio --name=hdd-qualification --filename=/srv/hdd-emu/qualification.bin --size=128m --rw=randread --bs=4k --ioengine=libaio --iodepth=1 --numjobs=1 --direct=1 --time_based=1 --runtime=30 --ramp_time=3 --group_reporting=1 --output-format=json`. Result: 117.01 IOPS, p50 7.96 ms, p99 16.58 ms. Uptime at start: 3:35; load average: 0.04, 0.02, 0.07. The IOPS and median latency match the expected HDD profile.
kayg commented 2026-10-02 10:27:46 +00:00
Author
Owner
Copy link

Results

Profile Result
Tab switches for Calendar, Files, Photos, Mail, Analytics, Money and Ask; warm/cold; Chromium/WebKit; 390/1440 px No timing samples. The profile stopped before browser measurement because the server did not become ready after seeding/indexing the Home.

The only log evidence was slow SQLite setup work during startup: VACUUM INTO 5.85 s, journal_mode = WAL 7.16 s, and CREATE TABLE semantic_meta 2.04 s. The runner hit its readiness deadline. It did not report a server crash or request result, so this is not a measured product regression. JSON output was not created. VM load was 0.04/0.88/1.43 at start and 3.81/2.77/2.11 at stop.

Attempted command

bun bench/tab-switch.mjs --local --engines chromium,webkit --viewports 390,1440 --themes light --conditions warm,cold --runs 3 --cold-runs 3 --json /root/perf-rerun/output/tab-switch-dev.json --production-build /root/perf-rerun/base/apps/web/build, with TAB_SWITCH_BUILD_BIN=/root/perf-rerun/bin/dev-calternal-server, CALTERNAL_E2E_ASSET_OVERRIDE=1, and Playwright browsers at /opt/ms-playwright. It was wrapped in /root/hdd-emu.sh run-limited under flock -w 14400 /root/perf.lock. I used the local cold-cache adapter and served the saved c4 production bundle. The matrix did not begin.

Decision: I used the light theme only because this request specified Tabs, engines, widths and warm/cold conditions, but no theme matrix. No product files changed.

## Results | Profile | Result | | --- | --- | | Tab switches for Calendar, Files, Photos, Mail, Analytics, Money and Ask; warm/cold; Chromium/WebKit; 390/1440 px | No timing samples. The profile stopped before browser measurement because the server did not become ready after seeding/indexing the Home. | The only log evidence was slow SQLite setup work during startup: `VACUUM INTO` 5.85 s, `journal_mode = WAL` 7.16 s, and `CREATE TABLE semantic_meta` 2.04 s. The runner hit its readiness deadline. It did not report a server crash or request result, so this is not a measured product regression. JSON output was not created. VM load was 0.04/0.88/1.43 at start and 3.81/2.77/2.11 at stop. ## Attempted command `bun bench/tab-switch.mjs --local --engines chromium,webkit --viewports 390,1440 --themes light --conditions warm,cold --runs 3 --cold-runs 3 --json /root/perf-rerun/output/tab-switch-dev.json --production-build /root/perf-rerun/base/apps/web/build`, with `TAB_SWITCH_BUILD_BIN=/root/perf-rerun/bin/dev-calternal-server`, `CALTERNAL_E2E_ASSET_OVERRIDE=1`, and Playwright browsers at `/opt/ms-playwright`. It was wrapped in `/root/hdd-emu.sh run-limited` under `flock -w 14400 /root/perf.lock`. I used the local cold-cache adapter and served the saved c4 production bundle. The matrix did not begin. Decision: I used the light theme only because this request specified Tabs, engines, widths and warm/cold conditions, but no theme matrix. No product files changed.
kayg commented 2026-10-02 10:30:31 +00:00
Author
Owner
Copy link

Final report

This was a measurement-only run. No tracked files changed and no commit was created. Branch job/perf-rerun is at c4a61e8cf090170f35b1bed3350d9de20c83ecd5, matching origin/dev after the required fetch/merge.

Issue Result
#573 Three baseline attempts and two complete feature runs. Listing, query and 10k sync beat the 5 s feature budget; 50-client move visibility did not. RSS and peak-CPU signals above 10% are tracked in #711 and #712. Full table and commands are on #573.
#632 Three interleaved A/B runs completed. Large Note readiness and keydown-to-frame improved; pointer burst changed +2.0%. No change crossed 10%. Full table and command are on #632.
#651 Three interleaved A/B runs completed. Reminder p95, CPU and RSS changes stayed within 10%. Full table and command are on #651.
#549 No tab-switch timing samples. The server missed its readiness deadline after slow SQLite setup during fixture preparation. Full attempt and command are on this issue.
#476 WebDAV PUT matrix not measured before the cutoff. Planned matrix command is on #476.
#641 Settings and Mail Blaze matrices not measured before the cutoff. Planned commands are on #641.

The #549, #632, #573 and #651 reports include the measurement commands, build SHAs, run counts and load records. The HDD emulation fio qualification and the failed #573 B2 setup are also recorded on this issue.

Files: none tracked. Temporary release binaries and web production builds were used for measurements and removed from the worktree. Head SHA: c4a61e8cf090170f35b1bed3350d9de20c83ecd5.

Gates: not run; there were no source changes to format, lint or test. Cleanup output: Removed 7833 files, 3.1GiB total.

Known gaps: no #549 timing samples; #476 and #641 remain unmeasured. The #573 benchmark reports peak sampled CPU rather than average CPU-seconds; #712 tracks that gap.

Decisions: I ran the #549 attempt in light theme because its requested matrix named Tabs, engines, widths and warm/cold states but no theme. The attempt failed before the matrix started. No product or design decisions were made.

UX gaps closed/left: not applicable; no UI changed.

## Final report This was a measurement-only run. No tracked files changed and no commit was created. Branch `job/perf-rerun` is at `c4a61e8cf090170f35b1bed3350d9de20c83ecd5`, matching `origin/dev` after the required fetch/merge. | Issue | Result | | --- | --- | | #573 | Three baseline attempts and two complete feature runs. Listing, query and 10k sync beat the 5 s feature budget; 50-client move visibility did not. RSS and peak-CPU signals above 10% are tracked in #711 and #712. Full table and commands are on #573. | | #632 | Three interleaved A/B runs completed. Large Note readiness and keydown-to-frame improved; pointer burst changed +2.0%. No change crossed 10%. Full table and command are on #632. | | #651 | Three interleaved A/B runs completed. Reminder p95, CPU and RSS changes stayed within 10%. Full table and command are on #651. | | #549 | No tab-switch timing samples. The server missed its readiness deadline after slow SQLite setup during fixture preparation. Full attempt and command are on this issue. | | #476 | WebDAV PUT matrix not measured before the cutoff. Planned matrix command is on #476. | | #641 | Settings and Mail Blaze matrices not measured before the cutoff. Planned commands are on #641. | The #549, #632, #573 and #651 reports include the measurement commands, build SHAs, run counts and load records. The HDD emulation fio qualification and the failed #573 B2 setup are also recorded on this issue. **Files:** none tracked. Temporary release binaries and web production builds were used for measurements and removed from the worktree. **Head SHA:** `c4a61e8cf090170f35b1bed3350d9de20c83ecd5`. **Gates:** not run; there were no source changes to format, lint or test. Cleanup output: `Removed 7833 files, 3.1GiB total`. **Known gaps:** no #549 timing samples; #476 and #641 remain unmeasured. The #573 benchmark reports peak sampled CPU rather than average CPU-seconds; #712 tracks that gap. **Decisions:** I ran the #549 attempt in light theme because its requested matrix named Tabs, engines, widths and warm/cold states but no theme. The attempt failed before the matrix started. No product or design decisions were made. **UX gaps closed/left:** not applicable; no UI changed.
kayg commented 2026-10-02 10:48:07 +00:00
Author
Owner
Copy link

Starting #549 on job/hddsql-549, based on c4a61e8cf090170f35b1bed3350d9de20c83ecd5 (origin/dev at worktree start). I will reproduce startup on the locked HDD-emulation perf VM, trace SQLite statements over 50 ms, and record fixes plus the corrected tab-switch matrix.

Starting #549 on `job/hddsql-549`, based on `c4a61e8cf090170f35b1bed3350d9de20c83ecd5` (`origin/dev` at worktree start). I will reproduce startup on the locked HDD-emulation perf VM, trace SQLite statements over 50 ms, and record fixes plus the corrected tab-switch matrix.
kayg commented 2026-10-02 13:06:05 +00:00
Author
Owner
Copy link

Full-fixture HDD reproduction finding (branch job/hddsql-549, head b7c9e36f9): the emulated filesystem qualifies at QD1 125.008 IOPS, p50 7.963 ms, p99 8.225 ms. The seeded-Home restart missed the 300 s readiness deadline. With CALTERNAL_SQLITE_SLOW_STATEMENT_MS=50, the initial fresh start logged WAL setup at 156–160 ms, VACUUM INTO at 161 ms and first semantic schema creation at 83 ms. A second run against the full seeded Home logged VACUUM INTO at 737 ms, semantic LSH inserts at 87–131 ms, and a semantic document insert at 51 ms. None explain a 300 s delay. That run also logged repeated Tantivy commits and SQLx pool acquisitions over 2 s, while /readyz stayed unavailable; the exact pre-listener phase is still unknown. I am adding opt-in startup phase timing before changing behavior.

Full-fixture HDD reproduction finding (branch `job/hddsql-549`, head `b7c9e36f9`): the emulated filesystem qualifies at QD1 125.008 IOPS, p50 7.963 ms, p99 8.225 ms. The seeded-Home restart missed the 300 s readiness deadline. With `CALTERNAL_SQLITE_SLOW_STATEMENT_MS=50`, the initial fresh start logged WAL setup at 156–160 ms, `VACUUM INTO` at 161 ms and first semantic schema creation at 83 ms. A second run against the full seeded Home logged `VACUUM INTO` at 737 ms, semantic LSH inserts at 87–131 ms, and a semantic document insert at 51 ms. None explain a 300 s delay. That run also logged repeated Tantivy commits and SQLx pool acquisitions over 2 s, while `/readyz` stayed unavailable; the exact pre-listener phase is still unknown. I am adding opt-in startup phase timing before changing behavior.
kayg commented 2026-10-02 13:17:40 +00:00
Author
Owner
Copy link

perf-arch-io (#663) reviewed the startup code at origin/dev c4a61e8cf and queued merge-round-7a 2f4482ded. It found three relevant startup phases, without running a new measurement:

  • Every restart awaits a full Index VACUUM snapshot before checking migrations. Focused follow-up #748 preserves #23's pre-upgrade safety but skips the unnecessary unchanged-restart copy. Your reported 737 ms VACUUM does not explain the 300 s timeout.
  • Recursive watcher registration is synchronous on Linux: notify 8.2.0 (the Cargo.lock version), inotify.rs:400 uses WalkDir and :547 waits on rx.recv. Search indexer.rs:504/:2597 and collaboration session.rs:703 each wait for their own full users-tree registration before bind. wire.rs:5538 starts a third Home watch in a Tokio task. #806 owns watcher reuse/registration scheduling; instrument these phases before attributing the timeout to them.
  • Semantic startup opens its schema before bind and sync_homes queues Home upserts before bind (wire.rs:2924), which can request the queued lazy model too. Whole-Home Notes/Photos work also follows Files completion. Independent SQLx/background workers do not inherit the later startup thread's idle IO class.

Post-bind contention is separate: Files full reconcile holds Root's shared mutation lock for an entire Home, including changed-file hashing (#750). Pure WAL reads can work while writes wait. Notes duplicate scans and Photos whole-library writer publication have detailed evidence on existing #704/#683; repeated unchanged PDF/text work is on #695.

No new number or exact cause of the readiness timeout is claimed. No product change, push or deploy. Please reuse these source checks with your opt-in startup phase timing; this audit does not duplicate that instrumentation.

perf-arch-io (#663) reviewed the startup code at origin/dev c4a61e8cf and queued merge-round-7a 2f4482ded. It found three relevant startup phases, without running a new measurement: - Every restart awaits a full Index VACUUM snapshot before checking migrations. Focused follow-up #748 preserves #23's pre-upgrade safety but skips the unnecessary unchanged-restart copy. Your reported 737 ms VACUUM does not explain the 300 s timeout. - Recursive watcher registration is synchronous on Linux: notify 8.2.0 (the Cargo.lock version), inotify.rs:400 uses WalkDir and :547 waits on rx.recv. Search indexer.rs:504/:2597 and collaboration session.rs:703 each wait for their own full users-tree registration before bind. wire.rs:5538 starts a third Home watch in a Tokio task. #806 owns watcher reuse/registration scheduling; instrument these phases before attributing the timeout to them. - Semantic startup opens its schema before bind and sync_homes queues Home upserts before bind (wire.rs:2924), which can request the queued lazy model too. Whole-Home Notes/Photos work also follows Files completion. Independent SQLx/background workers do not inherit the later startup thread's idle IO class. Post-bind contention is separate: Files full reconcile holds Root's shared mutation lock for an entire Home, including changed-file hashing (#750). Pure WAL reads can work while writes wait. Notes duplicate scans and Photos whole-library writer publication have detailed evidence on existing #704/#683; repeated unchanged PDF/text work is on #695. No new number or exact cause of the readiness timeout is claimed. No product change, push or deploy. Please reuse these source checks with your opt-in startup phase timing; this audit does not duplicate that instrumentation.
kayg commented 2026-10-02 15:16:59 +00:00
Author
Owner
Copy link

Perf VM finding (#549), captured while holding /root/perf.lock; VM load average inside the lock was 8.94 7.83 4.07.

The retained full-fixture startup log has one SQLite statement above 50 ms: VACUUM INTO ? at 737 ms. It also records repeated SQLx pool-acquire waits from 2.0 to 3.59 seconds during startup. Those waits are not statements, so statement-only diagnostics did not explain the readiness miss. The phase-timing build is committed as 0e1bd9235; the next locked boot will attribute the wait to its pre-listener phase.

Perf VM finding (#549), captured while holding `/root/perf.lock`; VM load average inside the lock was `8.94 7.83 4.07`. The retained full-fixture startup log has one SQLite statement above 50 ms: `VACUUM INTO ?` at 737 ms. It also records repeated SQLx pool-acquire waits from 2.0 to 3.59 seconds during startup. Those waits are not statements, so statement-only diagnostics did not explain the readiness miss. The phase-timing build is committed as `0e1bd9235`; the next locked boot will attribute the wait to its pre-listener phase.
kayg commented 2026-10-02 16:12:10 +00:00
Author
Owner
Copy link

Checkpoint at the owner’s ~4-hour limit. Issue #549 remains open; the startup fix and performance matrix are not complete.

Built and committed

  • 5f160df8d adds opt-in SQLx slow-statement logging via CALTERNAL_SQLITE_SLOW_STATEMENT_MS, including the DB and semantic SQLite pools.
  • b7c9e36f9 makes the benchmark retain complete redacted startup logs on readiness failure.
  • 0e1bd9235 adds opt-in timings for awaited pre-listener phases via CALTERNAL_STARTUP_PHASE_TIMING=1.

Files: Cargo.lock, bench/tab-switch.mjs, crates/calternal-db/{Cargo.toml,src/db.rs,src/lib.rs,src/sqlite.rs}, crates/calternal-embed/{Cargo.toml,src/store.rs,src/user_store.rs}, and crates/calternal-server/src/wire.rs.

Perf evidence

While holding /root/perf.lock, the retained full-fixture log showed VM load 8.94 7.83 4.07, one SQL statement over 50 ms (VACUUM INTO ?, 737 ms), and repeated SQLx pool-acquire waits from 2.0 to 3.59 seconds. These waits do not identify the slow pre-listener phase. This was a log review, not a new boot or tab-switch measurement.

Gate output

  • cargo fmt --check: exit 0, no output.
  • cargo clippy -p calternal-db --all-targets -- -D warnings: passed; Finished dev profile ... in 23.54s.
  • cargo test -p calternal-db: 11 unit and 16 integration tests passed; 1 ignored; doc tests passed.
  • cargo clippy -p calternal-embed --all-targets -- -D warnings: passed; Finished dev profile ... in 4m 01s.
  • cargo test -p calternal-embed: 31 passed, 4 ignored; doc tests passed.
  • node --check bench/tab-switch.mjs: exit 0, no output.
  • bun test bench/tab-switch.test.mjs:
bun test v1.4.2 (744846f84)

bench/tab-switch.test.mjs:
(pass) pending Calendar selection cannot paint the previous Photos route [0.17ms]
(pass) request timeout diagnostics redact credential headers, including ANSI lines [0.58ms]

 2 pass
 0 fail
 15 expect() calls
Ran 2 tests across 1 file. [136.00ms]
  • cargo clippy -p calternal-server --all-targets -- -D warnings (retry after a shared sccache disconnect):
Finished dev profile [unoptimized + debuginfo] target(s) in 34m 58s
  • cargo test -p calternal-server was stopped with exit 143 after 65 minutes; the last output was Compiling foreign-types v0.3.2. No test results were produced.
  • cargo build --release -p calternal-server was stopped with exit 143 during server linking under high shared-host load. No updated release binary was produced.
  • cargo clean: Removed 19118 files, 6.9GiB total. The generated apps/web/build output was removed.

Known gaps

No startup behavior was changed. The 737 ms statement and pool-acquire waits are not yet attributed to a phase. The instrumented release boot, root-cause fix, warm/cold tab-switch matrix, and latency-threshold check remain. I did not fetch/merge origin/dev or run an adversarial round; no API or route changed.

Decision

The design does not specify startup diagnostics. I kept both diagnostic flags opt-in, and the benchmark stores redacted logs under ignored target/tmp; SQL bind values are not logged.

Head: 0e1bd923528a8fa3ba1b8012c19e99ace46ef1ee.

Checkpoint at the owner’s ~4-hour limit. Issue #549 remains open; the startup fix and performance matrix are not complete. **Built and committed** - `5f160df8d` adds opt-in SQLx slow-statement logging via `CALTERNAL_SQLITE_SLOW_STATEMENT_MS`, including the DB and semantic SQLite pools. - `b7c9e36f9` makes the benchmark retain complete redacted startup logs on readiness failure. - `0e1bd9235` adds opt-in timings for awaited pre-listener phases via `CALTERNAL_STARTUP_PHASE_TIMING=1`. Files: `Cargo.lock`, `bench/tab-switch.mjs`, `crates/calternal-db/{Cargo.toml,src/db.rs,src/lib.rs,src/sqlite.rs}`, `crates/calternal-embed/{Cargo.toml,src/store.rs,src/user_store.rs}`, and `crates/calternal-server/src/wire.rs`. **Perf evidence** While holding `/root/perf.lock`, the retained full-fixture log showed VM load `8.94 7.83 4.07`, one SQL statement over 50 ms (`VACUUM INTO ?`, 737 ms), and repeated SQLx pool-acquire waits from 2.0 to 3.59 seconds. These waits do not identify the slow pre-listener phase. This was a log review, not a new boot or tab-switch measurement. **Gate output** - `cargo fmt --check`: exit 0, no output. - `cargo clippy -p calternal-db --all-targets -- -D warnings`: passed; `Finished dev profile ... in 23.54s`. - `cargo test -p calternal-db`: 11 unit and 16 integration tests passed; 1 ignored; doc tests passed. - `cargo clippy -p calternal-embed --all-targets -- -D warnings`: passed; `Finished dev profile ... in 4m 01s`. - `cargo test -p calternal-embed`: 31 passed, 4 ignored; doc tests passed. - `node --check bench/tab-switch.mjs`: exit 0, no output. - `bun test bench/tab-switch.test.mjs`: ```text bun test v1.4.2 (744846f84) bench/tab-switch.test.mjs: (pass) pending Calendar selection cannot paint the previous Photos route [0.17ms] (pass) request timeout diagnostics redact credential headers, including ANSI lines [0.58ms] 2 pass 0 fail 15 expect() calls Ran 2 tests across 1 file. [136.00ms] ``` - `cargo clippy -p calternal-server --all-targets -- -D warnings` (retry after a shared `sccache` disconnect): ```text Finished dev profile [unoptimized + debuginfo] target(s) in 34m 58s ``` - `cargo test -p calternal-server` was stopped with exit 143 after 65 minutes; the last output was `Compiling foreign-types v0.3.2`. No test results were produced. - `cargo build --release -p calternal-server` was stopped with exit 143 during server linking under high shared-host load. No updated release binary was produced. - `cargo clean`: `Removed 19118 files, 6.9GiB total`. The generated `apps/web/build` output was removed. **Known gaps** No startup behavior was changed. The 737 ms statement and pool-acquire waits are not yet attributed to a phase. The instrumented release boot, root-cause fix, warm/cold tab-switch matrix, and latency-threshold check remain. I did not fetch/merge `origin/dev` or run an adversarial round; no API or route changed. **Decision** The design does not specify startup diagnostics. I kept both diagnostic flags opt-in, and the benchmark stores redacted logs under ignored `target/tmp`; SQL bind values are not logged. Head: `0e1bd923528a8fa3ba1b8012c19e99ace46ef1ee`.
kayg commented 2026-10-02 16:20:47 +00:00
Author
Owner
Copy link

perfguards-impl / #798 verification finding (2026-10-02)

One local run reused bench/tab-switch.mjs and the shared release binary. The perf VM lock was occupied. The command requested Chromium, 1440 px, light, five warm Calendar → Files samples, macOS platform emulation, and the existing large Home fixture.

Setup failed before any route sample:

Error: Photos indexed 0 of 50000 fixture Items before the timeout
    at waitForPhotoIndex (.../bench/tab-switch.mjs:1158:12)
    at async main (.../bench/tab-switch.mjs:1375:26)

The build host load average was above 100 during the run. This is not a measured route regression and is not a PASS. I did not rerun. The initial legacy harness did not save a report for a setup failure. #798 now fixes that harness gap: setup and later failures save an incomplete report, preserve collected samples, redact credentials, and exit 2. Four harness tests pass, including this case. The populated profile setup remains unverified. Please investigate the shared-build/Home/projection compatibility in the existing #549 profile before accepting timing numbers.

perfguards-impl / #798 verification finding (2026-10-02) One local run reused `bench/tab-switch.mjs` and the shared release binary. The perf VM lock was occupied. The command requested Chromium, 1440 px, light, five warm Calendar → Files samples, macOS platform emulation, and the existing large Home fixture. Setup failed before any route sample: ``` Error: Photos indexed 0 of 50000 fixture Items before the timeout at waitForPhotoIndex (.../bench/tab-switch.mjs:1158:12) at async main (.../bench/tab-switch.mjs:1375:26) ``` The build host load average was above 100 during the run. This is not a measured route regression and is not a PASS. I did not rerun. The initial legacy harness did not save a report for a setup failure. #798 now fixes that harness gap: setup and later failures save an incomplete report, preserve collected samples, redact credentials, and exit 2. Four harness tests pass, including this case. The populated profile setup remains unverified. Please investigate the shared-build/Home/projection compatibility in the existing #549 profile before accepting timing numbers.
kayg commented 2026-10-02 17:27:52 +00:00
Author
Owner
Copy link

Independent read-only review of git diff origin/dev...0e1bd9235 on branch job/hddsql-549.

Findings

  • P2 — Startup timing skips pre-listener work. crates/calternal-server/src/wire.rs:1115 runs Root::recover() before the first timed phase. wire.rs:1267-1269 and wire.rs:1280 also await database operations without startup_phase. A pool-acquire wait in these operations has no phase name or elapsed time. Time all pre-listener work, including synchronous Root setup and each awaited database operation, and add a regression check for the phase logs.
  • P2 — Early-exit log capture can miss final output. bench/tab-switch.mjs:342-344 saves the output as soon as child.exitCode is set. Child stdio can remain open until the later close event, so final startup diagnostics can arrive after the snapshot. Wait for close with a bounded fallback before saving and add a harness regression for a final line written immediately before exit.

Both findings concern #549. Search for CALTERNAL_STARTUP_PHASE_TIMING returned this issue; no separate issue was filed.

No build, test, server, or browser command ran, as required by the read-only job. The report is committed at 85691d7781b16575a7b58a423865c2f87a677272 in review-hddsql-549.md; evidence is in audit-findings.md.

Independent read-only review of `git diff origin/dev...0e1bd9235` on branch `job/hddsql-549`. ## Findings - **P2 — Startup timing skips pre-listener work.** `crates/calternal-server/src/wire.rs:1115` runs `Root::recover()` before the first timed phase. `wire.rs:1267-1269` and `wire.rs:1280` also await database operations without `startup_phase`. A pool-acquire wait in these operations has no phase name or elapsed time. Time all pre-listener work, including synchronous Root setup and each awaited database operation, and add a regression check for the phase logs. - **P2 — Early-exit log capture can miss final output.** `bench/tab-switch.mjs:342-344` saves the output as soon as `child.exitCode` is set. Child stdio can remain open until the later `close` event, so final startup diagnostics can arrive after the snapshot. Wait for `close` with a bounded fallback before saving and add a harness regression for a final line written immediately before exit. Both findings concern #549. Search for `CALTERNAL_STARTUP_PHASE_TIMING` returned this issue; no separate issue was filed. No build, test, server, or browser command ran, as required by the read-only job. The report is committed at `85691d7781b16575a7b58a423865c2f87a677272` in `review-hddsql-549.md`; evidence is in `audit-findings.md`.
kayg commented 2026-10-02 19:44:44 +00:00
Author
Owner
Copy link

Starting follow-up on job/hddsql-549 at 0e1bd923528a8fa3ba1b8012c19e99ace46ef1ee, based on origin/dev c4a61e8cf090170f35b1bed3350d9de20c83ecd5. I will close the two independent-review findings, then trace the pool-acquire waits and move the startup VACUUM INTO work to idle background work. I will merge origin/dev once before final gates and report the measured run and gate output.

Starting follow-up on `job/hddsql-549` at `0e1bd923528a8fa3ba1b8012c19e99ace46ef1ee`, based on `origin/dev` `c4a61e8cf090170f35b1bed3350d9de20c83ecd5`. I will close the two independent-review findings, then trace the pool-acquire waits and move the startup VACUUM INTO work to idle background work. I will merge origin/dev once before final gates and report the measured run and gate output.
kayg commented 2026-10-02 20:10:00 +00:00
Author
Owner
Copy link

Finding: Node can set exitCode before child stdout/stderr close, so the harness could save a startup log without its final diagnostic. Commit 6e84a3bbf waits for stdio close with a 1 s bound in early-exit and already-stopped paths. A regression simulates buffered output arriving after the exit state and verifies the saved text includes it. node --check bench/tab-switch.mjs and bun test bench/tab-switch.test.mjs pass (3 tests).

Finding: Node can set `exitCode` before child stdout/stderr close, so the harness could save a startup log without its final diagnostic. Commit `6e84a3bbf` waits for stdio close with a 1 s bound in early-exit and already-stopped paths. A regression simulates buffered output arriving after the exit state and verifies the saved text includes it. `node --check bench/tab-switch.mjs` and `bun test bench/tab-switch.test.mjs` pass (3 tests).
kayg commented 2026-10-02 21:48:26 +00:00
Author
Owner
Copy link

Source audit finding (not yet confirmed by the perf boot): Db::connect sets the Index writer pool to one connection and its reader pool to eight. Worker::run enforces concurrency per job kind but has no total in-flight limit; start_background registers the system handlers plus every enabled Plugin kind and starts that Worker before build_live_app returns, so due jobs can contend for those shared pools before the listener binds. I am adding a server-specific global background-job bound and deferring Worker startup until after listener bind. The build host is currently saturated with concurrent server compiles; I will still run the requested perf profile under /root/perf.lock and label any load-limited results.

Source audit finding (not yet confirmed by the perf boot): `Db::connect` sets the Index writer pool to one connection and its reader pool to eight. `Worker::run` enforces concurrency per job kind but has no total in-flight limit; `start_background` registers the system handlers plus every enabled Plugin kind and starts that Worker before `build_live_app` returns, so due jobs can contend for those shared pools before the listener binds. I am adding a server-specific global background-job bound and deferring Worker startup until after listener bind. The build host is currently saturated with concurrent server compiles; I will still run the requested perf profile under `/root/perf.lock` and label any load-limited results.
kayg commented 2026-10-02 22:16:44 +00:00
Author
Owner
Copy link

Snapshot lead: Db::apply_migration_sets_with_snapshot now checks pending migration versions and checksums before creating a pre-upgrade image. Unchanged restarts skip VACUUM INTO; pending migrations keep the recovery snapshot required by DESIGN §27. Checksum drift and the no-snapshot restart path have regression coverage. Commit: dc5f9cdfaaf7fc5c4419e2b1aa2504dafba2229f.

Gate evidence for calternal-db:

  • cargo clippy -p calternal-db --all-targets -- -D warnings: Finished dev profile [unoptimized + debuginfo] target(s) in 2m 44s
  • cargo test -p calternal-db: test result: ok. 15 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 24.80s; queue integration tests: test result: ok. 16 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 2.15s; doc-tests: 0 passed; 0 failed.

Next I’m validating the bounded worker and first-API idle gate. No performance numbers are claimed yet.

Snapshot lead: `Db::apply_migration_sets_with_snapshot` now checks pending migration versions and checksums before creating a pre-upgrade image. Unchanged restarts skip `VACUUM INTO`; pending migrations keep the recovery snapshot required by DESIGN §27. Checksum drift and the no-snapshot restart path have regression coverage. Commit: `dc5f9cdfaaf7fc5c4419e2b1aa2504dafba2229f`. Gate evidence for `calternal-db`: - `cargo clippy -p calternal-db --all-targets -- -D warnings`: `Finished `dev` profile [unoptimized + debuginfo] target(s) in 2m 44s` - `cargo test -p calternal-db`: `test result: ok. 15 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 24.80s`; queue integration tests: `test result: ok. 16 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 2.15s`; doc-tests: `0 passed; 0 failed`. Next I’m validating the bounded worker and first-API idle gate. No performance numbers are claimed yet.
kayg commented 2026-10-02 22:49:30 +00:00
Author
Owner
Copy link

Startup pool contention implementation is committed as fbb35d35b0bbaae6c9699996fdaa9d8d16f04835. The server now waits until initial /api/ traffic completes and stays idle for two seconds (15-second post-bind fallback), then starts reconciliation and background work. The server Worker admits one handler across all registered kinds. Pending-migration snapshots remain before schema writes; unchanged restarts skip VACUUM INTO.

Verification: cargo fmt --check exited 0 with no output. The calternal-db clippy and test gates passed (reported in the prior comment). The first server clippy compile found a watch sender/receiver ordering error; that was corrected. The corrected clippy retry then sat in folio_wait_bit_common for over four minutes, and cargo test -p calternal-server entered the same I/O wait within 34 seconds. Both were stopped. At the time, 13 host processes were in uninterruptible I/O, with multiple other server clippy/test builds active. I am not claiming the server gates passed.

I have not run the HDD profile yet. No UI or route contract changed.

Startup pool contention implementation is committed as `fbb35d35b0bbaae6c9699996fdaa9d8d16f04835`. The server now waits until initial `/api/` traffic completes and stays idle for two seconds (15-second post-bind fallback), then starts reconciliation and background work. The server Worker admits one handler across all registered kinds. Pending-migration snapshots remain before schema writes; unchanged restarts skip `VACUUM INTO`. Verification: `cargo fmt --check` exited 0 with no output. The `calternal-db` clippy and test gates passed (reported in the prior comment). The first server clippy compile found a watch sender/receiver ordering error; that was corrected. The corrected clippy retry then sat in `folio_wait_bit_common` for over four minutes, and `cargo test -p calternal-server` entered the same I/O wait within 34 seconds. Both were stopped. At the time, 13 host processes were in uninterruptible I/O, with multiple other server clippy/test builds active. I am not claiming the server gates passed. I have not run the HDD profile yet. No UI or route contract changed.
kayg commented 2026-10-02 22:57:15 +00:00
Author
Owner
Copy link

Completed

Head: fbb35d35b0bbaae6c9699996fdaa9d8d16f04835.

  • Added named timing for synchronous Root recovery and the previously unwrapped database setup calls.
  • Fixed tab-switch startup-log capture to wait for child stdio close, with a bounded fallback and regression coverage.
  • Added migration preflight: unchanged restarts skip VACUUM INTO; a pending migration still gets its required pre-migration snapshot.
  • Deferred startup reconciliation and queued work until initial /api/ traffic is complete and idle for two seconds. A 15-second post-bind fallback keeps scheduled work running when no browser opens the app.
  • Added a shared one-handler limit across server Worker kinds.

Files: bench/tab-switch.mjs, bench/tab-switch.test.mjs, crates/calternal-db/src/migrations.rs, crates/calternal-db/src/worker.rs, crates/calternal-server/src/main.rs, crates/calternal-server/src/wire.rs.

Commits: 6e84a3bbf, 49706398c, dc5f9cdfa, fbb35d35b. No route contract or UI changed. UX gaps: not applicable.

Gates

cargo fmt --check exited 0 and printed no output.

cargo clippy -p calternal-db --all-targets -- -D warnings:

    Finished `dev` profile [unoptimized + debuginfo] target(s) in 2m 44s

cargo test -p calternal-db:

test result: ok. 15 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 24.80s
test result: ok. 16 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 2.15s
test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s

The first server clippy attempt caught a watch sender/receiver ordering mistake; it is fixed. The corrected cargo clippy -p calternal-server --all-targets -- -D warnings child then stayed in folio_wait_bit_common for over four minutes with no output and was stopped. cargo test -p calternal-server entered the same uninterruptible I/O wait within 34 seconds and was stopped. At the time, 13 host processes were in uninterruptible I/O and several other server builds were active. These server gates are incomplete, not passing.

Cleanup output:

     Removed 14418 files, 5.8GiB total

Performance and remaining work

I did not build the release binary or run the HDD profile because the build host remained under heavy I/O load. No before/after measurements are claimed. For the merge round, rerun the server gates and build once:

export CARGO_PROFILE_DEV_DEBUG=line-tables-only CARGO_INCREMENTAL=0 CARGO_BUILD_JOBS=4 TMPDIR="$PWD/target/tmp"
cargo clippy -p calternal-server --all-targets -- -D warnings
cargo test -p calternal-server
cargo build -p calternal-server --release

Then alternate one cold run per binary three times. bench/tab-switch.mjs records startup first paint and tab-switch timings, server CPU/RSS, and load average; it acquires the perf VM lock for its phases. Set TAB_SWITCH_VM_LOCK_WAIT_SECONDS=14400 on every invocation:

for run in 1 2 3; do
  TAB_SWITCH_VM_LOCK_WAIT_SECONDS=14400 TAB_SWITCH_BUILD_BIN=/mnt/hdd/targets/release-shared/release/calternal-server bun bench/tab-switch.mjs --runs 1 --cold-runs 1 --conditions cold --engines chromium --viewports 1440 --themes light --tabs calendar,photos --json "target/tmp/hddsql-549-before-${run}.json"
  TAB_SWITCH_VM_LOCK_WAIT_SECONDS=14400 TAB_SWITCH_BUILD_BIN=/mnt/hdd/targets/jobs/hddsql-549/release/calternal-server bun bench/tab-switch.mjs --runs 1 --cold-runs 1 --conditions cold --engines chromium --viewports 1440 --themes light --tabs calendar,photos --json "target/tmp/hddsql-549-after-${run}.json"
done

Decisions and gaps

  • The exact idle thresholds were not set in DESIGN. I chose a two-second quiet window after the first API burst and a 15-second fallback after bind.
  • DESIGN §27 requires a consistent pre-migration snapshot. That remains on the rare pending-migration path; normal unchanged restarts do not run VACUUM INTO.
  • Remaining gaps: complete the server gates and the six alternating HDD measurements in the merge round. No other UX gaps were introduced.
## Completed Head: `fbb35d35b0bbaae6c9699996fdaa9d8d16f04835`. - Added named timing for synchronous Root recovery and the previously unwrapped database setup calls. - Fixed tab-switch startup-log capture to wait for child stdio close, with a bounded fallback and regression coverage. - Added migration preflight: unchanged restarts skip `VACUUM INTO`; a pending migration still gets its required pre-migration snapshot. - Deferred startup reconciliation and queued work until initial `/api/` traffic is complete and idle for two seconds. A 15-second post-bind fallback keeps scheduled work running when no browser opens the app. - Added a shared one-handler limit across server Worker kinds. Files: `bench/tab-switch.mjs`, `bench/tab-switch.test.mjs`, `crates/calternal-db/src/migrations.rs`, `crates/calternal-db/src/worker.rs`, `crates/calternal-server/src/main.rs`, `crates/calternal-server/src/wire.rs`. Commits: `6e84a3bbf`, `49706398c`, `dc5f9cdfa`, `fbb35d35b`. No route contract or UI changed. UX gaps: not applicable. ## Gates `cargo fmt --check` exited 0 and printed no output. `cargo clippy -p calternal-db --all-targets -- -D warnings`: ```text Finished `dev` profile [unoptimized + debuginfo] target(s) in 2m 44s ``` `cargo test -p calternal-db`: ```text test result: ok. 15 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 24.80s test result: ok. 16 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 2.15s test result: ok. 0 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.00s ``` The first server clippy attempt caught a watch sender/receiver ordering mistake; it is fixed. The corrected `cargo clippy -p calternal-server --all-targets -- -D warnings` child then stayed in `folio_wait_bit_common` for over four minutes with no output and was stopped. `cargo test -p calternal-server` entered the same uninterruptible I/O wait within 34 seconds and was stopped. At the time, 13 host processes were in uninterruptible I/O and several other server builds were active. These server gates are incomplete, not passing. Cleanup output: ```text Removed 14418 files, 5.8GiB total ``` ## Performance and remaining work I did not build the release binary or run the HDD profile because the build host remained under heavy I/O load. No before/after measurements are claimed. For the merge round, rerun the server gates and build once: ```sh export CARGO_PROFILE_DEV_DEBUG=line-tables-only CARGO_INCREMENTAL=0 CARGO_BUILD_JOBS=4 TMPDIR="$PWD/target/tmp" cargo clippy -p calternal-server --all-targets -- -D warnings cargo test -p calternal-server cargo build -p calternal-server --release ``` Then alternate one cold run per binary three times. `bench/tab-switch.mjs` records startup first paint and tab-switch timings, server CPU/RSS, and load average; it acquires the perf VM lock for its phases. Set `TAB_SWITCH_VM_LOCK_WAIT_SECONDS=14400` on every invocation: ```sh for run in 1 2 3; do TAB_SWITCH_VM_LOCK_WAIT_SECONDS=14400 TAB_SWITCH_BUILD_BIN=/mnt/hdd/targets/release-shared/release/calternal-server bun bench/tab-switch.mjs --runs 1 --cold-runs 1 --conditions cold --engines chromium --viewports 1440 --themes light --tabs calendar,photos --json "target/tmp/hddsql-549-before-${run}.json" TAB_SWITCH_VM_LOCK_WAIT_SECONDS=14400 TAB_SWITCH_BUILD_BIN=/mnt/hdd/targets/jobs/hddsql-549/release/calternal-server bun bench/tab-switch.mjs --runs 1 --cold-runs 1 --conditions cold --engines chromium --viewports 1440 --themes light --tabs calendar,photos --json "target/tmp/hddsql-549-after-${run}.json" done ``` ## Decisions and gaps - The exact idle thresholds were not set in DESIGN. I chose a two-second quiet window after the first API burst and a 15-second fallback after bind. - DESIGN §27 requires a consistent pre-migration snapshot. That remains on the rare pending-migration path; normal unchanged restarts do not run `VACUUM INTO`. - Remaining gaps: complete the server gates and the six alternating HDD measurements in the merge round. No other UX gaps were introduced.
kayg commented 2026-10-03 01:46:16 +00:00
Author
Owner
Copy link

Crate gates on job/hddsql-549 (server, db, embed), run by Claude:

  • cargo fmt --check: exit 0
  • clippy -D warnings: calternal-server, calternal-db, calternal-embed all exit 0
  • cargo test -p calternal-db: test result: ok. 15 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 5.52s / test result: ok. 16 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 0.17s
  • cargo test -p calternal-embed: test result: ok. 31 passed; 0 failed; 4 ignored; 0 measured; 0 filtered out; finished in 0.29s
  • cargo test -p calternal-server, first run (real, fails alone 3/3): 
    thread 'wire::tests::live_apps_run_in_separate_processes' (2298135) panicked at crates/calternal-server/src/wire.rs:7013:13:
thread 'wire::tests::first_start_snapshot_contains_the_pre_migration_index' (2298148) has overflowed its stack
fatal runtime error: stack overflow, aborting
test result: FAILED. 108 passed; 1 failed; 3 ignored; 0 measured; 0 filtered out; finished in 4.84s
    Two causes found:
    1. All three live-app children overflow the 2 MiB debug test-thread stack (the deferred-startup change made the build_live_app future deeper). They pass with 4 MiB. Production polls it on the 8 MiB main thread, so this is test-only.
    2. With more stack, full_app_setup_session_config_and_backup panicked at wire.rs:7858 (dedup scrub poll timed out after 60 s). The test dropped LiveApp, which closed the new startup gate, so start_background never ran.
  • Fix: 5dd850804 test(server): run live-app children with the server stack and release startup work (#549) — child processes get RUST_MIN_STACK=SERVER_WORKER_STACK_SIZE; the test releases startup_work as server_main does. No assertion changed.
  • Re-run: 
    test result: ok. 109 passed; 0 failed; 3 ignored; 0 measured; 0 filtered out; finished in 7.62s

Ready for the merge round (crate gates). Note for review: any other code that builds a LiveApp must release startup_work, or background jobs never start.

Crate gates on `job/hddsql-549` (server, db, embed), run by Claude: - `cargo fmt --check`: exit 0 - clippy `-D warnings`: calternal-server, calternal-db, calternal-embed all exit 0 - `cargo test -p calternal-db`: `test result: ok. 15 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 5.52s` / `test result: ok. 16 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 0.17s` - `cargo test -p calternal-embed`: `test result: ok. 31 passed; 0 failed; 4 ignored; 0 measured; 0 filtered out; finished in 0.29s` - `cargo test -p calternal-server`, first run (real, fails alone 3/3): ``` thread 'wire::tests::live_apps_run_in_separate_processes' (2298135) panicked at crates/calternal-server/src/wire.rs:7013:13: thread 'wire::tests::first_start_snapshot_contains_the_pre_migration_index' (2298148) has overflowed its stack fatal runtime error: stack overflow, aborting test result: FAILED. 108 passed; 1 failed; 3 ignored; 0 measured; 0 filtered out; finished in 4.84s ``` Two causes found: 1. All three live-app children overflow the 2 MiB debug test-thread stack (the deferred-startup change made the `build_live_app` future deeper). They pass with 4 MiB. Production polls it on the 8 MiB main thread, so this is test-only. 2. With more stack, `full_app_setup_session_config_and_backup` panicked at wire.rs:7858 (dedup scrub poll timed out after 60 s). The test dropped `LiveApp`, which closed the new startup gate, so `start_background` never ran. - Fix: 5dd850804 `test(server): run live-app children with the server stack and release startup work (#549)` — child processes get `RUST_MIN_STACK=SERVER_WORKER_STACK_SIZE`; the test releases `startup_work` as `server_main` does. No assertion changed. - Re-run: ``` test result: ok. 109 passed; 0 failed; 3 ignored; 0 measured; 0 filtered out; finished in 7.62s ``` Ready for the merge round (crate gates). Note for review: any other code that builds a `LiveApp` must release `startup_work`, or background jobs never start.
kayg commented 2026-10-06 13:43:06 +00:00
Author
Owner
Copy link

Performance review results from #1124 are recorded in docs/perf/2026-10-06-1124.md.

Path Result Evidence / limitation
Files folder, 5,000 entries Render marker p50/p95 14,912 / 15,358 ms at 4× CPU throttle Virtual list kept 27 DOM rows; renderer marker was 13,442 ms. No Chrome trace was captured. Warm HDD route setup hit its 12-minute limit before producing a sample.
Calendar Week, 121 provider Events Zoom 8.5 fps; frame p50/p95 100.1 / 216.7 ms Trace has 1,060.5 ms UpdateLayoutTree, 530.1 ms Layout, and 42 long tasks (236 ms max). The per-frame --hour update likely drives grid layout; this is not isolated to one source function.
Photos grid scroll 13.7 fps baseline; candidate 27.1 fps Interleaved A/B: frame p50/p95 improved from 66.7 / 100.1 ms to 33.3 / 50.1 ms. Scroll-scoped backdrop-filter suppression is in 685b1105f.

Files and Calendar measurements used the native /dev/sda1 filesystem and are diagnostic, not HDD-qualified. Photos A/B also ran on native storage. The post-change Photos HDD run could not start because another job held /root/perf.lock; no HDD screenshot set was produced. The likely Files derived-view and Calendar layout costs remain follow-ups.

Performance review results from #1124 are recorded in `docs/perf/2026-10-06-1124.md`. | Path | Result | Evidence / limitation | | --- | --- | --- | | Files folder, 5,000 entries | Render marker p50/p95 14,912 / 15,358 ms at 4× CPU throttle | Virtual list kept 27 DOM rows; renderer marker was 13,442 ms. No Chrome trace was captured. Warm HDD route setup hit its 12-minute limit before producing a sample. | | Calendar Week, 121 provider Events | Zoom 8.5 fps; frame p50/p95 100.1 / 216.7 ms | Trace has 1,060.5 ms `UpdateLayoutTree`, 530.1 ms `Layout`, and 42 long tasks (236 ms max). The per-frame `--hour` update likely drives grid layout; this is not isolated to one source function. | | Photos grid scroll | 13.7 fps baseline; candidate 27.1 fps | Interleaved A/B: frame p50/p95 improved from 66.7 / 100.1 ms to 33.3 / 50.1 ms. Scroll-scoped backdrop-filter suppression is in `685b1105f`. | Files and Calendar measurements used the native `/dev/sda1` filesystem and are diagnostic, not HDD-qualified. Photos A/B also ran on native storage. The post-change Photos HDD run could not start because another job held `/root/perf.lock`; no HDD screenshot set was produced. The likely Files derived-view and Calendar layout costs remain follow-ups.
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
dev
wip/draw3-1101
wip/draw2-1101
wip/draw-1101
wip/directory-1199-r
wip/directory-1199
wip/delete-1119
wip/collabrev-1197
wip/collabloss-1197
wip/cards2-1083
wip/cards-1083
wip/canvas-visual
wip/canvasvis2-976
wip/calhdr-1112
wip/calcards-1115
wip/browserfix
wip/blocks-1125
wip/allday-1107
wip/agenda-decks
wip/agenda-1086
wip/adv7c-1105
wip/txentry-1198
wip/trayicons2-1095
wip/trayicons-1095
wip/tagperf-1186
wip/sidebar3-1094
wip/rev2-webperf
wip/rev2-money-ident
job/adv-1202
wip/restyle-notes
wip/previewcard-1098
wip/palette2-1123
wip/palette-1093
wip/onboard2-1141
wip/onboard-1141.aborted-early
wip/onboard-1141
wip/nlpchip-1127
wip/morph-1104
wip/merge-round-7c5
wip/merge-round-7c4
job/notifloop-1194
wip/merge-round-7c3
wip/merge-round-7c2
wip/merge-round-7c
wip/mchrome-1084
wip/mailghost2-1094
wip/mailghost-1094
wip/kbpreview2-1118
wip/kbpreview-1118
wip/kanban-1092
wip/importhang-1121
wip/hiderev-1153
wip/hide4-1153
wip/hide3-1153
wip/hide2-1153
wip/hide-1153
wip/editreg-1132
job/collabloss-1197
wip/editorrail3-1113
wip/editorrail2-1113
wip/editorrail-1113
wip/e2e-b2-1071
wip/e2e-b-1071
wip/draw4-1101
job/onboard-1141
job/hide-1153
job/perf-1124
job/perf2-1124
job/tocrail-1191
job/restyle-settings
wip/restyle-settings
job/segmented-1200
wip/notifloop-1194
job/tagperf-1186
wip/segmented-1200
job/restyle-files
job/tagdnd-1187
job/merge30
job/cards-1179
wip/cards2-1179
wip/cards-1179
wip/tocrail-1191
wip/tagdnd-1187
wip/restyle-files
wip/perf-1124
wip/merge30j
job/restyle-notes
job/wizchoices-1140
wip/wizchoices-1140
wip/restyle-1190
job/moneyfmt-1180
job/txentry-1198
wip/moneyfmt2-1180
wip/moneyfmt-1180-r
wip/moneyfmt-1180
job/pillglass-1189
job/flags-1181
wip/flags-1181
job/restyle-1190
job/restyle-mailmoney
job/restyle-search
job/settingsreg-1195
job/wizard-1140
site/website
wip/wizardrev2-1140
wip/wizardrev-1140
wip/wizard5-1140
wip/wizard4-1140
wip/wizard3-1140
wip/wizard2-1140
wip/wizard-1140
wip/pillglass-1189
wip/settingsreg-1195
job/merge29
job/fu-1171
wip/merge29j
wip/fu-1171
job/fu-1166
job/directory-1199
job/proflog-1204
job/txresearch-1188
wip/fu-1166
job/merge28
job/search-1066
wip/search-1066
wip/merge28j
job/gateslot-1182
job/bulkimport-1157
job/mailnet-1160
wip/mailnetrev-1160
wip/mailnet-1160
wip/bulkrev-1157
wip/bulkimport-1157
job/startup-1161
wip/startup-1161
job/merge27
job/linkcards-1151
wip/linkcards3-1151
wip/linkcards2-1151
wip/linkcards-1151
job/traydate-1144
wip/traydate3-1144
wip/traydate2-1144
wip/traydate-1144
job/draw-1101
wip/merge27j
job/blockpill-1152
wip/blockpill3-1152
wip/blockpill2-1152
wip/blockpill-1152
job/minihover-1149
wip/minihover2-1149
wip/minihover-1149
job/merge25
wip/merge25-r
wip/merge25b
wip/merge25
job/inspector-1129
job/tags-1110
wip/inspector3-1129
wip/inspector2-1129
wip/inspector-1129
wip/tagsrev-1110
wip/tags2-1110
wip/tags-1110
job/dates-1148
wip/datesrev-1148
wip/dates2-1148
wip/dates-1148
job/licence-1145
wip/licence2-1145
wip/licence-1145
job/selfhost-1156
job/merge23
wip/merge23
job/tagfilter-1109
wip/tagfilter2-1109
wip/tagfilter-1109
job/kbd-1134
wip/kbd2-1134
wip/kbd-1134
job/palfoot-1137
wip/selfhost-1156
wip/palfoot2-1137
wip/palfoot-1137
job/toggle-1158
wip/toggle-1158
job/kbpreview-1118
job/docratchet-1155
job/perflint-1133
job/devtests-1159
wip/docratchet-1155
wip/devtests-1159
job/segv-1136
wip/toast-1142
wip/segv-1136
job/toast-1142
job/blockreload-1147
wip/blockreload-1147
job/font-1150
wip/font-1150
job/importui-1120
job/minimonth-1149
wip/importui-1120
wip/minimonth-1149
job/depcheck-1146
wip/perflint-1133
wip/depcheck-1146
job/calcards-1115
job/blocks-1125
job/plus-1128
job/shift-1138
wip/plus2-1128
wip/plus-1128
wip/shift-1138
job/moneyfid-1130
job/editorrail-1113
wip/moneyrev-1130
wip/moneyfid-1130
job/noext-851
wip/noext-851
wip/noext3-851
wip/noext2-851
job/week-1135
wip/week-1135
job/editreg-1132
job/smoke-1122
wip/smoke-1122
job/docs-1143
job/palette2-1123
job/calhdr-1112
job/nlpchip-1127
job/mailghost-1094
job/reconnect-1131
wip/reconnect-1131
job/trayicons-1095
job/delete-1119
job/importhang-1121
job/cards-1083
job/palette-1093
job/mchrome-1084
job/e2e-a-1071
job/canvas-visual
job/previewcard-1098
job/allday-1107
wip/e2e-a2-1071
wip/e2e-a-1071
job/e2e-b-1071
job/adv7c-1105
job/kanban-1092
job/agenda-1086
job/merge-round-7c
job/morph-1104
wip/surfaces-p2
job/merge-round-9
wip/merge-round-9
job/7cfix-small
wip/7cfix-small
job/mailui-1078
job/merge-round-8
wip/merge-round-8
wip/mailui-1078
job/mailround-1038
job/applemail-accept
wip/settitle-1068
wip/mailround2-1038
wip/mailround-1038
wip/e2e-7b
job/crash-1069
wip/crash-1069
job/searchlost-1066
wip/searchlost-1066
job/7b-reconcile
job/flake-1065
wip/flake-1065
wip/merge-round-7b7
wip/merge-round-7b6
wip/merge-round-7b5
wip/merge-round-7b4
wip/7b-reconcile
job/appupdate-1059
job/nfd-1044
wip/appupdate-1059
job/e2e-7b
job/loop-1062
wip/loop-1062
job/pdfprev-1045
job/invtoggle-1053
wip/pdfprev-1045
wip/nfd-1044
wip/invtoggle-1053
job/7bfix-e2e
job/mailstress-b
wip/7bfix-e2e
wip/mailstress-b
job/7bfix-adv
wip/7bfix-adv
job/mailstress-a
job/stack-1054
wip/stack-1054
wip/mailstress-a
job/mailstress-1038
wip/mailstress-1038
job/upload500-1051
wip/upload500-1051
job/share-1034
wip/share-1034
job/syncerr-1037
job/7bfix-photos
wip/7bfix-photos
job/paste-1036
job/setside-1039
wip/setside-1039
wip/paste-1036
job/lease-1042
wip/syncerr-1037
wip/lease-1042
job/7bfix-data
job/passkeybind-1043
wip/apprevoke-1041
job/invite-1035
wip/invite-1035
job/merge-round-7b2
wip/merge-round-7b2
job/mailproxy-486
job/apprevoke-1041
job/rebuild-1033
job/pillborder-1029
wip/pillborder-1029
wip/mailproxy-486
wip/applemail-486
job/headless-998
wip/headless-998
job/groups-1028
wip/groups-1028
job/rebuildwarn-1016
wip/rebuildwarn-1016
job/startup-1011
wip/startup-1011
job/monthpill-1009
job/bgthumb-1025
job/sharetitle-1012
wip/monthpill-1009
wip/bgthumb-1025
wip/sharetitle-1012
job/canvas-cards-977
wip/canvas-cards-977
job/canvas-pencil-978
job/canvas-sketch-990
wip/canvas-sketch-990
wip/canvas-pencil-978
job/canvas-files-989
wip/canvas-files-989
job/canvas-collab-991
wip/canvas-collab-991
job/weekscroll-1018
wip/weekscroll-1018
wip/canvas-core-976
job/canvas-core-976
job/round-drag
wip/round-drag
job/round-settings
job/browserfix
wip/oapi-974
job/oapi-974
job/hist2-integrate
job/mailhtml-726
wip/mailhtml-726
wip/hist2-integrate
job/moneyfu-984
job/drag-1015
wip/drag-1015
job/rename-1017
wip/rename-1017
job/hist2-api
wip/hist2-api
job/oneacct-1014
wip/oneacct-1014
wip/moneyfu-984
job/hist2-bench
job/hist2-restore
wip/hist2-bench
job/hist2-write
job/hotfix-724
wip/hotfix-724
wip/hist2-write
wip/hist2-restore
job/hist2-store
job/hist2-ui
wip/hist2-ui
wip/hist2-store
job/searchstarve-965
job/shutdown-963
wip/shutdown-963
wip/pubedit-981
job/pubedit-981
job/analytics-973
wip/searchstarve-965
job/authflash-850
job/weeklane-969
job/pvtitle-1004
job/hist-975
wip/authflash-850
job/voicepill-617
wip/pvtitle-1004
job/headring-1003
wip/weeklane-969
wip/voicepill-617
wip/headring-1003
wip/analytics-973
job/agentscope-980
wip/thumbsandbox-988
job/thumbsandbox-988
wip/hist-975
job/links-856
wip/links-856
job/davetag-966
wip/davetag-966
job/filesstorm-1000
job/hoverpad-725
wip/filesstorm-1000
job/ffmpegblas-993
job/merge-round-7a
wip/hoverpad-725
wip/ffmpegblas-993
job/nowdot-1002
wip/verify-7a
job/noteid-857
wip/nowdot-1002
wip/noteid-857
wip/merge-round-7a
wip/agentscope-980
job/imapedge
job/a11yfix2
wip/imapedge-941
wip/imapedge
wip/a11yfix2
job/notetask-986
job/logheading
wip/logheading-998
job/textthumb-652
job/photolive-987
wip/photolive-987
job/davactive-983
job/savefix-985
job/tabicons-607
wip/davactive-983
wip/tabicons-607
wip/notetask-986
wip/savefix-985
job/dirid-627
job/buildspeed-1007
wip/dirid-627
job/agenda-decks
job/perfguards-impl
job/undo-a11y
wip/undo-a11y
job/mailperf
job/wal-824
wip/settings-50
job/settings-50
job/notesfilter-606
wip/notesfilter-606
job/surfaces-p2
wip/wal-824
job/maillayouts
wip/mailperf
wip/maillayouts
job/taskmeta-659
job/money-ident
wip/money-ident
wip/taskmeta-659
job/errstates
wip/perfguards-impl
job/headings-881
wip/headings-881
wip/errstates
job/voice-619
job/gaps-827
job/notesperf
wip/notesperf
wip/voice-619
job/hddsql-549
job/perf-stream-668
wip/perf-stream-668
wip/deeplinks-fix
job/deeplinks-fix
job/authfix
job/docsfix-rust
wip/docsfix-rust
job/webperf
job/docsfix-web
job/datafix2
job/webdav-lock-476
job/copyfix
wip/copyfix
wip/webperf
job/focus-658
wip/protofix
job/mediafix
job/protofix
wip/mediafix
job/agentfix
job/hhmm-724
wip/agentfix
job/undo-722
job/reuse
wip/webdav-lock-476
wip/reuse
job/scopefix
job/datafix
wip/hhmm-724
wip/undo-722
job/surfaces-p1
wip/hddsql-549
job/voicememos-618
wip/datafix2
wip/surfaces-p1
job/fix-940
wip/fix-940
job/blaze-surfaces
wip/datafix
wip/blaze-surfaces
job/taskday-655
job/linknav-639
wip/linknav-639
wip/gaps-827
job/isolation-707
job/audiophotos-720
wip/audiophotos-720
job/advfind-664
wip/voicememos-618
wip/taskday-655
wip/isolation-707
wip/advfind-664
wip/scopefix
wip/focus-658
job/testgaps
wip/testgaps
job/overscroll-718
wip/authfix
job/deps
wip/overscroll-718
job/rev2-agentfix
job/rev2-money-ident
job/rev2-mailperf
wip/deps
job/hardening-728
wip/hardening-728
job/searchgen-832
wip/searchgen-832
job/photopw-849
job/mailsql-825
wip/photopw-849
job/sharefix
wip/sharefix
job/rev2-mailhtml-726
job/rev2-perfguards
job/copyval-723
job/lightglass-r2
wip/lightglass-r2
wip/docsfix-web
job/copy-audit
job/macinterop-staging-r2
job/design-sync
job/rev2-taskmeta-659
job/rev2-webperf
job/docs-audit
job/rev2-advfind-664
job/rev2-mailproxy-486
job/states-audit
job/rev2-datafix
job/design-drift
job/test-gaps
job/rev2-voicememos-618
job/rev2-mediafix
job/rev2-deps
job/rev2-datafix2
job/licence-audit
job/issue-hygiene
job/rev2-protofix
job/rev2-voice-619
job/rev2-isolation-707
job/rev2-surfaces-p1
job/deeplink-audit2
job/rev2-audiophotos-720
wip/test-gaps
job/rev2-overscroll-718
job/rev2-undo-722
wip/states-audit
job/rev2-dropmd-719
job/rev2-linknav-639
job/merge-7b-plan
wip/merge-7b-plan
job/rev2-taskday-655
wip/mailsql-825
job/rev2-webdav-lock-476
job/rev2-browserfix
wip/design-drift
job/rev2-hddsql-549
wip/deeplink-audit2
job/rev2-scopefix
job/rev2-authfix
job/rev2-hardening-728
job/rev2-wal-824
job/rev2-sharefix
job/calsidebar-638
job/chrome-audit
job/ioperf
wip/ioperf
wip/chrome-audit
wip/calsidebar-638
job/dropmd-719
wip/dropmd-719
job/ocr-build
wip/ocr-build
job/blaze-settings
wip/copyval-723
job/toastring-721
wip/toastring-721
job/deployfix-732
wip/deployfix-732
wip/blaze-settings
job/money-import-recheck
job/rev-a11y
job/perf-arch-db
job/rev-7b-data
wip/textthumb-652
wip/perf-arch-db
job/sec-protocols
job/sidehdr-660
job/rev-7b-security
job/research-surfaces
job/rev-design-gaps
job/rev-mcp-api
wip/sidehdr-660
job/perf-arch-memory
wip/sec-protocols
job/perf-arch-bundle
job/snapedge-714
wip/rev-mcp-api
job/sec-supplychain
wip/research-surfaces
job/perf-arch-sync
job/rev-consistency
job/perf-arch-server
wip/perf-arch-server
wip/perf-arch-memory
job/perf-arch-io
job/perf-arch-client
job/sec-fs
job/sec-mcp-scopes
job/sec-sharing
job/perf-guards
job/sec-browser
job/sec-admin-deploy
job/sec-auth
wip/snapedge-714
job/bgpicker-717
wip/perf-arch-bundle
wip/money-import-recheck
job/advsetup-654
wip/bgpicker-717
wip/advsetup-654
job/burst-709
job/kbdcaps-710
job/app-pw-chooser
wip/burst-709
wip/app-pw-chooser
job/imaptest-625
wip/kbdcaps-710
job/fix-499
wip/fix-499
job/perf-mut-667
job/calimg-589
job/perf-snap-666
wip/calimg-589
wip/perf-snap-666
wip/perf-mut-667
job/perf-cache-665
wip/perf-cache-665
job/voicefiles-620
wip/voicefiles-620
job/admin-burst-705
wip/admin-burst-705
job/voicememos-review
wip/voicememos-review
wip/ryw-653
job/ryw-653
job/writeonopen-661
job/instant-663
wip/writeonopen-661
job/money-import-review
wip/money-import-review
wip/importjs-610
review/integrations-407-round6
wip/integrations-review
job/dragghost-612
wip/dragghost-612
job/integrations
wip/integrations
job/decider-656
job/merge-round-6
job/perf-rerun
wip/merge-round-6
job/integrations-review-round5
job/selalign-576
wip/selalign-576
job/mcp-events-491
job/files-631
job/cal-e2e-569
wip/cal-e2e-569
job/reload-423
wip/reload-423
wip/mcp-events-491
wip/files-631
job/notesbridge-644
wip/notesbridge-644
job/editor-series
job/calcard-series
wip/calcard-series
job/mcp-events-review-491
wip/mcp-events-review
wip/editor-series
job/quirks-546
job/integrations-recheck
job/tocrail-636
wip/tocrail-636
wip/quirks-546
wip/reminders-643
job/reminders-643
wip/davscale-573
job/davscale-573
job/integrations-review
wip/ocr-eval-584
job/ocr-eval-584
job/esc-537
wip/esc-537
job/toastname-586
wip/toastname-586
job/submenu-579
wip/submenu-579
job/tasks-mode
wip/tasks-mode
job/agentdocs-630
job/dupwrite-634
wip/agentdocs-630
wip/dupwrite-634
job/lightglass-588
wip/lightglass-588
job/tabswitch-549
job/ghosttask-623
wip/ghosttask-623
job/toaststack-616
job/weekstate-609
job/mailsync-613
wip/mailsync-613
wip/weekstate-609
job/maildup-626
wip/tabswitch-549
wip/maildup-626
wip/toaststack-616
job/motion-611
wip/motion-611
job/tlstest-601
wip/tlstest-601
job/perf-495
job/floating-sheet
wip/floating-sheet
job/remdup-585
wip/remdup-585
job/fix-502
wip/fix-502
job/attachplay-622
job/perf-batch
wip/perf-batch-563
wip/perf-495
hotfix/mail-sync-diag
job/mail-m3
wip/mail-m3
job/attach-poof-603
job/calhover-608
job/editorbar-604
job/mentions-605
job/merge-round-4
job/allday-514
wip/merge-round-4
wip/allday-514
job/merge-round-4a
wip/merge-round-4a
job/sharestack-580
job/fix-501
wip/sharestack-580
wip/fix-501
job/perf-batch-563
job/apw-cache-review
wip/apw-cache-review
job/probe-520
wip/probe-520
job/mac-393
wip/mac-393
job/header-571
job/flake-513
wip/flake-513
job/docs-thumb-547
wip/header-571
job/webcal-572
wip/webcal-572
wip/shortcuts-542
job/shortcuts-542
wip/docs-thumb-547
job/caldav-stress
wip/caldav-stress
wip/sweep-478
job/apw-cache-512
wip/apw-cache-512
job/money-empty-540
wip/restart-505
wip/money-empty-540
wip/fix-510
job/restart-505
job/fix-503
job/perf-496
wip/perf-496
job/fix-498
wip/fix-498
job/info-inspector-465
wip/info-inspector-465
job/fix-510
job/fix-507
wip/fix-507
wip/fix-503
job/fix-493
job/money-kinds
wip/money-kinds
job/hygiene-548
job/merge-round-3
wip/fix-493
job/drag-snap-536
wip/merge-round-3
wip/merge-round-0930
wip/drag-snap-536
job/align-538
wip/align-538
job/bg-flash
wip/bg-flash
job/money-import
job/search-count-544
wip/search-count-544
wip/money-import
job/settings-key-541
wip/settings-key-541
job/toast-539
job/preview-421
wip/preview-421
wip/toast-539
job/tasks-500-531
job/title-plain-526
wip/title-plain-526
wip/tasks-500-531
job/notes-bridge
wip/parity-484
job/parity-484
job/files-slow
job/crash-525
wip/notes-bridge
wip/files-slow
wip/crash-525
job/kbd-motion-527
wip/bg-422
job/analytics-504
wip/analytics-504
wip/kbd-motion-527
job/upload-pill-523
wip/upload-pill-523
wip/tray-order
job/tray-order
wip/overflow-mid
wip/merge-round-2
job/perf-494
wip/perf-494
wip/mcp-fast-492
wip/motion-477
wip/asr-ab-489
wip/theme-variants-506
wip/overflow-511
wip/week-header-508
wip/attach-427
job/dav-delete-471
job/iso-435
wip/iso-435
wip/files-sel-keys
wip/dav-delete-471
job/align-253
job/siwc-490
wip/siwc-490
job/money-kinds-review
wip/align-253
wip/money-kinds-review
job/small-bugs-3
wip/overlay-title-487
wip/multiget-500
wip/hidden-420
wip/webcal-ui
wip/webcal-431
job/perf-367
job/location
wip/small-bugs-3
wip/location
wip/perf-367
wip/admin-deny-483
job/tag-unicode-473
wip/tag-unicode-473
job/blur-436
wip/photos-470
wip/blur-436
wip/small-bugs-4
wip/hunt-20260930
wip/settings-hdr-482
wip/chips-416
job/dedup-375
wip/dedup-375
job/doc-stack
wip/doc-stack
job/tokens-literals
wip/tokens-literals
job/jobs-leftovers
wip/send-fast
wip/paste-467
wip/money-numbers
job/money-plugin
wip/money-plugin
job/break-dav
wip/merge-batch
wip/crossday-469
wip/mac-verify
wip/mail-m2
wip/break-dav
wip/money-review2
job/money-md
job/modes-424
wip/money-md
wip/jobs-leftovers
job/agenda-413
wip/agenda-413
wip/modes-424
job/recog-417
wip/recog-417
wip/bounce-425
wip/ab-384-luna
job/webdav-perf
wip/webdav-perf
job/toast-ring
wip/toast-ring
job/money-review
wip/money-review
wip/micro-motion
wip/settings-card
wip/minical
job/notes-imap-428
job/least-priv
wip/ui-small-2
wip/flaky-426
wip/drag-end-418
job/jank
wip/jank
wip/least-priv
wip/docs-site
job/agenda
job/sec-batch
wip/sec-batch
wip/per-user-index
job/area-calendars
wip/area-calendars
job/parity
wip/parity
job/documents-research
wip/documents-research
job/test-infra
job/reminders-sync
wip/small-bugs-2
wip/reminders-sync
wip/gestures
job/google-oauth
wip/tags-merge
wip/tags
job/e2e-theme
wip/e2e-theme
job/icon-align
wip/test-infra
wip/select-align
wip/editor-385
job/voice
wip/webdav
job/webdav
job/app-pw-ui
job/editor-integrity
wip/editor-integrity
wip/voice
wip/quota
wip/cal-followups
wip/icon-align
job/composer-scale
wip/composer-scale
job/jobs-page
wip/jobs-page
job/hig-type
wip/hig-type
wip/app-pw-ui
job/motion-spring
job/mcp
wip/motion-spring
wip/mcp
job/small-bugs
wip/push-hosts
job/profile-sign
wip/touch-369
wip/profile-sign
job/mobile-focus
wip/mobile-focus
wip/ui-polish-354
wip/small-bugs
wip/dup-task
job/toast-polish
job/app-pw-scopes
wip/toast-polish
wip/app-pw-scopes
wip/cli-agent
wip/selection-pills
job/preview-attach
wip/preview-attach
job/dav-proppatch
wip/dav-proppatch
wip/cal-switcher
job/atomic-race
wip/atomic-race
job/photos-shared
wip/photos-shared
wip/cal-grid
wip/note-rewrite
wip/search-rebuild
job/mail-m1
job/paperless-import
wip/paperless-import
wip/mail-m1
wip/hidden-activity
wip/search-d
wip/pricing-research
wip/cursors
wip/auto-scheme
job/single-pills
wip/single-pills
wip/xuser-matrix
wip/money-format
wip/app-pw-setup
wip/purge-dos
wip/vault-health
wip/caldav-apple
wip/xuser-audit
wip/e2e-green
wip/tabbar
wip/adv-harness
wip/maple-mono
job/search-fix
wip/search-fix
wip/search-perf-c
job/adv-harness
wip/sidebar-headers
job/glass
wip/temp-index
job/polish
wip/polish
wip/file-protocols
wip/money-research
wip/glass
wip/voice-models
wip/collab-redo
job/voice-research
wip/hunt-20260928
wip/notes-actions-research
wip/search-pad
wip/search-perf
wip/search-sticky
wip/editor-undo
wip/chrome-rules
wip/motion
wip/appearance-research
wip/appearance
wip/audit-bugs
wip/cal-glass
wip/block-actions
wip/authz-order
wip/event-stripes
wip/chrome-sidebar
wip/auth-flaky
wip/robust-2
wip/gate-fix
wip/menu-blur
wip/import-calternaljs
wip/tray-fix
job/import-calternaljs
wip/index-order
wip/audit-fixes
wip/search-chevrons
research/mail
wip/phone-chrome
wip/dedup-break
wip/csp
wip/ui-audit
wip/select-toast
wip/perf
wip/flat-layout
wip/fonts
wip/event-tint
wip/sync-converge
wip/data-split
wip/glass-audit
wip/robustness
wip/sync-chaos
wip/search-thumbs
wip/fuzz
wip/menu-icons
wip/search-pill
wip/sync-changing
wip/heading-links
wip/date-formats
wip/a11y
wip/break-editor
wip/e2e-fix
wip/settings-sections
wip/sync-root-guard
wip/search-palette
wip/share-edit
job/toasts
wip/toasts
wip/cont-analytics
wip/authz-review
wip/popovers
wip/overlay-glass
wip/change-feed
wip/editor-modes
wip/composer-align
wip/cont-agenda
wip/agenda-merge
job/agent-conventions
wip/agent-conventions
wip/backend-misc
job/route-audit
wip/route-audit
wip/ui-batch
wip/heif-hardening
wip/grid-resize
wip/ask-page
wip/webmcp
job/deeplink-audit
wip/deeplinks
wip/shortcuts
wip/cont-tz-days
main
No results found.
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
kayg/calternal#549
No description provided.