kayg/calternal
Watch
1
0
Fork
You've already forked calternal
 0
Code Issues 190 Pull requests 61 Projects Releases Packages Wiki Activity Actions 26

Merge train BM14c #897

Closed
kayg wants to merge 68 commits from train/2026-11-03-bm14c into dev
pull from: train/2026-11-03-bm14c
merge into: kayg:dev
Conversation 7 Commits 68 Files changed 43 +8202 -998
kayg commented 2026-11-03 13:35:21 +00:00
Owner
Copy link

BM14c, based on origin/dev 41f2efa387b74175971a838f602535089d0d5fb4. Head b59e4a5e2155b2b71c6246b98a4e20679073cf26. Do not merge.

PR Head merged (verified before merge) Outcome
#850 af6fb861 --no-ff, clean
#851 d94771b6 --no-ff, clean
#852 656b61bf --no-ff, clean
#829 97d790f6 --no-ff, clean
#835 3755691d (current head) --no-ff, clean
#895 3c1d8deb --no-ff, clean

No conflicts. First-run gates, full matrix and perf gate results follow in comments. No fixes, no retries.

BM14c, based on origin/dev `41f2efa387b74175971a838f602535089d0d5fb4`. Head `b59e4a5e2155b2b71c6246b98a4e20679073cf26`. Do not merge. | PR | Head merged (verified before merge) | Outcome | |---|---|---| | #850 | af6fb861 | --no-ff, clean | | #851 | d94771b6 | --no-ff, clean | | #852 | 656b61bf | --no-ff, clean | | #829 | 97d790f6 | --no-ff, clean | | #835 | 3755691d (current head) | --no-ff, clean | | #895 | 3c1d8deb | --no-ff, clean | No conflicts. First-run gates, full matrix and perf gate results follow in comments. No fixes, no retries.
kayg added 181 commits 2026-11-03 13:35:21 +00:00 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U5oTUycqAh7DT72mdSUztk
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U5oTUycqAh7DT72mdSUztk
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U5oTUycqAh7DT72mdSUztk
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U5oTUycqAh7DT72mdSUztk
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U5oTUycqAh7DT72mdSUztk
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01U5oTUycqAh7DT72mdSUztk
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The room under a phone's Style sheet loads with the idle screens
Some checks failed
CI / server (pull_request) Has been cancelled
Details
CI / rust (pull_request) Has been cancelled
Details
CI / web (pull_request) Has been cancelled
Details
Markdown extraction contract / markdown-native-contract (pull_request) Has been cancelled
Details
Markdown extraction contract / markdown-editor-contract (pull_request) Has been cancelled
Details
7a6d5adf4c
Presets: the thumbnail's page is not .page; backdrop, opacity and width specs follow the new Style sheet
Some checks failed
CI / server (pull_request) Has been cancelled
Details
CI / rust (pull_request) Has been cancelled
Details
CI / web (pull_request) Has been cancelled
Details
Markdown extraction contract / markdown-native-contract (pull_request) Has been cancelled
Details
Markdown extraction contract / markdown-editor-contract (pull_request) Has been cancelled
Details
c1ef1b78ed
Merge port/style-2
Some checks failed
Markdown extraction contract / markdown-native-contract (pull_request) Has been cancelled
Details
CI / server (pull_request) Has been cancelled
Details
CI / rust (pull_request) Has been cancelled
Details
CI / web (pull_request) Has been cancelled
Details
Markdown extraction contract / markdown-editor-contract (pull_request) Has been cancelled
Details
a0f425ac08 
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
dee3e961 replaced the fixed 450 ms x4 waits with still(); a settled page
draws no frames, so the screencast held 15-19 frames instead of 40+ and
'frames > 20' failed 16/20 unloaded and 20/20 under load. Wait for the
frame count (state), threshold unchanged.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Opening the note (5 s), typing 55 keys (6 s) and two palette round trips
(6 s) at 6x CPU took 18 s of a 30 s test on a quiet machine and hit the
limit whenever the machine was shared. Loading+typing and switching modes
are now separate tests; nothing about the timeout or the shift check changed.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
e2e: vim :noh waits for the marks to come back; lazy-ui marks the window; links waits for the job first
Some checks failed
CI / server (pull_request) Has been cancelled
Details
CI / rust (pull_request) Has been cancelled
Details
CI / web (pull_request) Has been cancelled
Details
Markdown extraction contract / markdown-native-contract (pull_request) Has been cancelled
Details
Markdown extraction contract / markdown-editor-contract (pull_request) Has been cancelled
Details
3c1d8deb19 
- vim.spec '/ searches...': Esc from a half-typed '?cher' removes the marks and puts the
  earlier search's three back about 60 ms later (probe: 0,3,3,...). :noh sent inside that gap
  clears nothing and the marks then return. Wait for the 3 before ':', and for the bar's
  field to hold the focus before typing. Flaked 4/20 on origin/dev, 1/20 on the BM14b head.
- lazy-ui.spec 'failed ... idle warming recovers': detect a reload with a mark on the window,
  not performance.timeOrigin (WebKit derives it from the wall clock; BM14b saw it move by 1 ms).
- links.spec 'Try again': wait for the server's job (15 s poll) before the 5 s URL check.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Merge #895 (3c1d8deb) into BM14c
Some checks failed
CI / server (pull_request) Has been cancelled
Details
CI / rust (pull_request) Has been cancelled
Details
CI / web (pull_request) Has been cancelled
Details
Markdown extraction contract / markdown-native-contract (pull_request) Has been cancelled
Details
Markdown extraction contract / markdown-editor-contract (pull_request) Has been cancelled
Details
b59e4a5e21
kayg commented 2026-11-03 14:06:03 +00:00
Author
Owner
Copy link

BM14c in progress. Head b59e4a5e2155b2b71c6246b98a4e20679073cf26, base origin/dev 41f2efa387b74175971a838f602535089d0d5fb4. Merged --no-ff in the requested order, every head matched the list (#835 current head 3755691d); no conflicts.

Train locks held on dev-01 and dev-02 (owner shells with EXIT trap), CI=1, SLOT_TRAIN=1, on-disk TMPDIR, --workers=1 --retries=0, df checked before each step (dev-01 ≥ 238 GiB, dev-02 ≥ 300 GiB free).

First-run gates so far (dev-01 unless noted), all pass: bun install --frozen-lockfile (both), cargo fmt --all --check, cargo clippy --workspace --all-targets --locked -j4 -- -D warnings, cargo test --workspace -j4, cargo build -j4 (both), bun run build (both), perf/web.sh --bundle-only --gate, ci/check-case-paths.mjs, ci/check-no-evidence.mjs.

Running now: the full Playwright matrix (Chromium on dev-01, WebKit on dev-02), then apps/web bun run check, then the perf-VM comparison against origin/dev.

BM14c in progress. Head `b59e4a5e2155b2b71c6246b98a4e20679073cf26`, base origin/dev `41f2efa387b74175971a838f602535089d0d5fb4`. Merged `--no-ff` in the requested order, every head matched the list (#835 current head `3755691d`); no conflicts. Train locks held on dev-01 and dev-02 (owner shells with EXIT trap), `CI=1`, `SLOT_TRAIN=1`, on-disk TMPDIR, `--workers=1 --retries=0`, df checked before each step (dev-01 ≥ 238 GiB, dev-02 ≥ 300 GiB free). First-run gates so far (dev-01 unless noted), all pass: `bun install --frozen-lockfile` (both), `cargo fmt --all --check`, `cargo clippy --workspace --all-targets --locked -j4 -- -D warnings`, `cargo test --workspace -j4`, `cargo build -j4` (both), `bun run build` (both), `perf/web.sh --bundle-only --gate`, `ci/check-case-paths.mjs`, `ci/check-no-evidence.mjs`. Running now: the full Playwright matrix (Chromium on dev-01, WebKit on dev-02), then `apps/web` `bun run check`, then the perf-VM comparison against origin/dev.
kayg commented 2026-11-03 15:43:38 +00:00
Author
Owner
Copy link

BM14c interim (head b59e4a5e2155b2b71c6246b98a4e20679073cf26). First-run matrix is RED: 17 failing cases on Chromium (dev-01) and 17 on WebKit (dev-02), same spec set on both browsers; passed 1381 (chromium half) + 1351 (webkit half). Not merging. Throttled projects (skipped by Playwright because their chromium dependencies failed) got their first attempts separately with --no-deps on the train binary: throttled-1440 3 pass, throttled-390 5 pass, 0 fail.

Failing cases on the train (first error; all deterministic in the first run):

  • cover.spec.ts:124, :143 (chromium-1440/390, webkit-1440/390): [data-testid=style-cover] Change button / "No cover" not found.
  • width.spec.ts:48 (1440 only, timeout), :86 (1440, tab "Wide" not found), :177 (1440 and 390, tab "Narrow" not found).
  • opacity.spec.ts:122 (1440/390): test timeout 30 s.
  • style-sheet.spec.ts:26, :49 (1440/390): test timeout 30 s.
  • note-bars.spec.ts:319 (1440, timeout) and :237 (390, "the Text page is shorter than Page": 394 vs < 245).
  • chromium-1440 only vim.spec.ts:387 (Saved not visible); webkit-1440 only accent-labels.spec.ts:26 (accent radios: 0 found, expected 7).

Lead (no fix made): svelte-check on the train reports src/components/style/StyleSheet.svelte:142: Property 'noteId' is missing in type '{}' but required for <PageTab />. The merge of #850/#851/#852 renders Page tab without its noteId prop, which fits the missing Width/Cover/Opacity rows. It also reports 2 type errors in e2e/review4-signout.spec.ts:225-226 (never; check against dev pending).

apps/web bun run check fails on both train and origin/dev 41f2efa3 with src/lib/actions.gen.ts is out of date with contracts/actions.json (pre-existing, not the train). Since that stops before svelte-check, svelte-check was run directly on the train (3 errors above); the same run on origin/dev is queued.

Next: one origin/dev comparison per failing case, --repeat-each=10 on the train, then the perf table (append, server hot paths, browser first paint are running on perf-01/02).

BM14c interim (head `b59e4a5e2155b2b71c6246b98a4e20679073cf26`). **First-run matrix is RED**: 17 failing cases on Chromium (dev-01) and 17 on WebKit (dev-02), same spec set on both browsers; passed 1381 (chromium half) + 1351 (webkit half). Not merging. Throttled projects (skipped by Playwright because their chromium dependencies failed) got their first attempts separately with `--no-deps` on the train binary: throttled-1440 3 pass, throttled-390 5 pass, 0 fail. Failing cases on the train (first error; all deterministic in the first run): - `cover.spec.ts:124`, `:143` (chromium-1440/390, webkit-1440/390): `[data-testid=style-cover]` Change button / "No cover" not found. - `width.spec.ts:48` (1440 only, timeout), `:86` (1440, tab "Wide" not found), `:177` (1440 and 390, tab "Narrow" not found). - `opacity.spec.ts:122` (1440/390): test timeout 30 s. - `style-sheet.spec.ts:26`, `:49` (1440/390): test timeout 30 s. - `note-bars.spec.ts:319` (1440, timeout) and `:237` (390, "the Text page is shorter than Page": 394 vs < 245). - chromium-1440 only `vim.spec.ts:387` (`Saved` not visible); webkit-1440 only `accent-labels.spec.ts:26` (accent radios: 0 found, expected 7). Lead (no fix made): `svelte-check` on the train reports `src/components/style/StyleSheet.svelte:142`: `Property 'noteId' is missing in type '{}' but required` for `<PageTab />`. The merge of #850/#851/#852 renders Page tab without its `noteId` prop, which fits the missing Width/Cover/Opacity rows. It also reports 2 type errors in `e2e/review4-signout.spec.ts:225-226` (`never`; check against dev pending). `apps/web` `bun run check` fails on both train and origin/dev `41f2efa3` with `src/lib/actions.gen.ts is out of date with contracts/actions.json` (pre-existing, not the train). Since that stops before svelte-check, svelte-check was run directly on the train (3 errors above); the same run on origin/dev is queued. Next: one origin/dev comparison per failing case, `--repeat-each=10` on the train, then the perf table (append, server hot paths, browser first paint are running on perf-01/02).
kayg commented 2026-11-03 16:19:14 +00:00
Author
Owner
Copy link

BM14c result: RED. Do not merge.

Head b59e4a5e2155b2b71c6246b98a4e20679073cf26, base origin/dev 41f2efa387b74175971a838f602535089d0d5fb4. Merged --no-ff in order, heads verified first: #850 af6fb861, #851 d94771b6, #852 656b61bf, #829 97d790f6, #835 3755691d (current), #895 3c1d8deb. No conflicts. CI=1, SLOT_TRAIN=1, train lock with EXIT trap on both VMs, on-disk TMPDIR, --workers=1 --retries=0, df before each step (dev-01 ≥ 235 GiB, dev-02 ≥ 300 GiB free). First run counted, nothing retried. No fixes made.

Gates (first run)

Gate Command Result
install (both VMs) bun install --frozen-lockfile pass
fmt cargo fmt --all --check pass
clippy cargo clippy --workspace --all-targets --locked -j4 -- -D warnings pass
Rust cargo test --workspace -j4 pass
cargo build (both) cargo build -j4 pass
web build (both) cd apps/web && bun run build pass
bundle budgets PERF_DIST=$PWD/apps/web/dist perf/web.sh --bundle-only --gate pass
case paths node ci/check-case-paths.mjs pass
evidence node ci/check-no-evidence.mjs pass
apps/web bun run check fails at actions.gen.ts is out of date with contracts/actions.json. The same failure happens on origin/dev 41f2efa3 (built the same way), so it is pre-existing, not the train.
svelte-check run directly (the check script stops before it) bunx svelte-check --tsconfig ./tsconfig.json train: 3 errors; dev: 2. The train-only error: src/components/style/StyleSheet.svelte:142:158 Property 'noteId' is missing in type '{}' but required (<PageTab />). The 2 in e2e/review4-signout.spec.ts:225-226 are on dev too.

Playwright matrix (every project, one attempt each, retry 0)

Chromium half on dev-01, WebKit half on dev-02, admin-setup on each, via the .mts wrapper as in BM14b.

VM / project Passed Failed Skipped
chromium / chromium-1440 433 10 625
chromium / chromium-390 892 7 169
chromium / throttled-1440 * 3 0 2
chromium / throttled-390 * 5 0 0
chromium / admin-chromium-1440, -390 12, 12 0 9, 9
chromium / admin-setup, flow, passkeys 1, 1, 1 0 0
chromium / prefs-chromium 7 0 0
chromium / look-chromium-1440, -390 5, 5 0 0
chromium / contrast 9 0 0
chromium / sidebar-band-chromium 3 0 0
webkit / webkit-1440 430 10 628
webkit / webkit-390 888 7 173
webkit / admin-webkit-1440, -390 9, 9 0 9, 9
webkit / admin-setup 1 0 0
webkit / prefs-webkit 3 0 4
webkit / look-webkit-1440, -390 4, 4 0 1, 1
webkit / sidebar-band-webkit 3 0 0
Total 2732 + 8* 34

* The throttled projects depend on the failed chromium projects, so Playwright skipped them. Their first attempts were run right after with --no-deps on the same train binary and dist.

Known flakes sidebar-edge.spec.ts:8, webkit-1440 editor.spec.ts:173, webkit-390 missing-links.spec.ts:245 did not occur.

Failures (34 first-run failures = 17 cases on each browser)

One comparison on origin/dev 41f2efa3 (built and run the same way, exact case, one attempt) and --repeat-each=10 --no-deps on the train. "pass n/10" is the train repeat.

Case Projects failed in matrix First error origin/dev Train repeat
cover.spec.ts:124 Style sheet Cover row all four (1440, 390, both browsers) [data-testid=style-cover] Change button not found spec does not exist on dev 0/10 on each of the four
cover.spec.ts:143 "Use this picture as the cover" all four style-cover "No cover" not found spec does not exist on dev 0/10 on each
style-sheet.spec.ts:26 all four test timeout 30 s spec does not exist on dev 0/10 on each
style-sheet.spec.ts:49 all four test timeout 30 s spec does not exist on dev 0/10 on each
note-bars.spec.ts:319 Style popover tabs 1440 chromium, webkit timeout 30 s spec does not exist on dev 0/10 on each
note-bars.spec.ts:237 Style tabs follow height 390 chromium, webkit "the Text page is shorter than Page": expected < 245, got 394 spec does not exist on dev chromium 0/10, webkit 1/10
opacity.spec.ts:122 (dev :120) slider to 0% all four timeout 30 s pass on all four 0/10 on each
width.spec.ts:48 1440 chromium, webkit timeout 30 s pass (both) 0/10 on each
width.spec.ts:86 1440 chromium, webkit tab "Wide" not found pass (both) 0/10 on each
width.spec.ts:177 all four tab "Narrow" not found (900px) pass on all four 0/10 on each
vim.spec.ts:387 (dev :380) : commands chromium-1440 Saved not visible fails on dev too (reproduces on baseline) 6/10 pass: flaky on both, not the train
accent-labels.spec.ts:26 webkit-1440 accent radios: 0 found, expected 7 pass 10/10 pass: first-run flake, not reproduced

So 10 distinct cases fail on every repeat only on the train (cover ×2, style-sheet ×2, note-bars ×2, opacity, width ×3). vim:380/387 fails on dev too; accent-labels:26 did not reproduce. Likely cause (not fixed): StyleSheet.svelte renders <PageTab /> without the required noteId after the #850/#851/#852 merge (svelte-check error above); the missing Width, Cover and Opacity rows are on that tab.

Perf (new gate: no regression vs origin/dev)

Train b59e4a5e against origin/dev 41f2efa3, release binaries built on dev-01 and relayed with matching SHA-256 on both perf VMs (head calternal 56787f47…, dev 2c6a9b93…, same calternal-perf 4631c52b…). Every run under ~/bin/perf-lock, seed 1, pairs alternating, perf VM loads quiet (0.3–5.2).

Server, perf-02 (calternal-perf --sizes 10k,100k --runs 5 --cold-runs 0 --idle-secs 0 --storm-edits 1, 3 alternating pairs; medians of per-run p50 / p95; spread = min–max of per-run p95 across the 3 runs):

Metric Home dev p50 / p95 head p50 / p95 p95 delta dev spread head spread Verdict
sync.round_trip (ms) 10k 0.34 / 0.42 0.49 / 0.64 +52% 0.39–0.46 0.63–0.65 regression (ranges do not overlap)
sync.round_trip (ms) 100k 0.35 / 0.45 0.57 / 0.71 +58% 0.42–0.50 0.68–0.81 regression
open.note (ms) 10k 0.91 / 1.84 1.28 / 2.06 +12% 1.46–2.22 1.79–2.17 within noise (p50 +0.37 ms, watch)
open.note (ms) 100k 5.64 / 12.3 6.04 / 10.6 −14% 11.2–13.4 8.45–13.1 flat
open.api_get (ms) 10k / 100k 2.94 / 10.6, 6.59 / 12.1 3.04 / 12.7, 6.16 / 13.5 +20% / +12% 6.3–11.6, 11.7–19.0 11.4–20.1, 9.6–16.5 within noise
home.page / page_title / page_next / page_tag (ms) 10k 0.88 / 1.78, 0.84 / 1.06, 0.94 / 1.27, 6.09 / 7.78 0.85 / 2.21, 0.68 / 0.99, 0.91 / 1.31, 5.75 / 7.92 +24%, −7%, +3%, +2% wide within noise
home.page / page_title / page_next / page_tag (ms) 100k 1.07 / 2.63, 0.85 / 1.71, 0.98 / 1.60, 57.9 / 65.3 1.23 / 1.83, 0.92 / 1.51, 1.08 / 1.37, 61.1 / 69.1 −30%, −11%, −15%, +6% page_tag 61.9–68.6 66.8–70.9 within noise; page_tag budget 50 ms is over on dev too
home.first_page (ms) 10k / 100k 2.86 / 5.32, 4.20 / 6.81 2.79 / 6.11, 5.35 / 9.19 +15% / +35% 3.1–6.8, 4.6–12.9 4.4–6.6, 6.5–12.1 within noise (single sample per restart)
search.query (ms, prefix/typo mix of the standard harness) 10k / 100k 1.94 / 4.56, 2.26 / 12.3 1.71 / 4.38, 2.32 / 5.94 −4% / −52% flat or better
search.first_query (ms) 10k / 100k 5.7 / 32.8, 8.9 / 52.0 5.4 / 24.8, 7.3 / 31.0 −24% / −40% flat or better
sync.save_ack (ms) 10k 2.70 / 3.15 2.93 / 4.98 +58% 3.0–3.3 3.25–7.5 one slow run; p50 +0.23 ms
sync.save_ack (ms) 100k 3.54 / 5.11 3.85 / 5.27 +3% within noise
sync.durable_write, journal.catch_up, startup, shutdown, mem.idle/peak both −5%…+4% (shutdown 10k +32%: 64 vs 84.5 ms, both runs bimodal 64/85 on dev too) flat

200 appends (30 alternating pairs, perf-01, CALTERNAL_NOTES_SAVE_MS=15, same append.py as #792): total time median dev 1.308 s, head 1.328 s = +1.50%; per-append p95 median 11.55 → 12.10 ms (+4.8%); pair ratios 0.79–1.21. Noise control, the same script with the dev binary against a copy of itself (30 pairs): +2.11%. So the +1.50% is inside run-to-run noise.

Browser, perf-01 (Chromium 4× CPU slowdown, 10k Home, perf/boot.perf.ts, list-search.perf.ts, blaze.perf.ts; 3 alternating runs per side; each side served by its own server binary and dist; WebKit not available on perf VMs):

Metric dev p50 / p95 head p50 / p95 p95 delta per-run medians dev / head Verdict
boot.cold_fcp (ms) 820 / 916 904 / 956 +4.4% (p50 +10%) 804–828 / 860–936 regression (separated)
boot.cold_note (first note painted, ms) 1051 / 1186 1118 / 1204 +1.5% (p50 +6.3%) 1017–1073 / 1113–1198 regression at p50 (separated), p95 within noise
boot.cold_list (ms) 1051 / 1186 1113 / 1199 +1.1% (p50 +5.9%) 1017–1073 / 1027–1180 within noise
boot.cold_tti / cold_cpu / cold_tbt 1893 / 2507, 1477 / 1732, 199 / 254 1980 / 2496, 1632 / 1743, 187 / 230 −0.4%, +0.6%, −9% within noise
boot.warm_note / warm_list / warm_tti (ms) 531 / 971, 357 / 707, 1496 / 1980 534 / 977, 355 / 717, 1480 / 1991 +0.6%, +1.4%, +0.5% flat
list_search.key (keystroke to paint, ms) 7.6 / 10.6 7.9 / 12.1 +14% (p50 +3.9%) 7.5–7.8 / 7.8–8.2 small regression (separated)
blaze.shown / back_shown / live (ms) 47.5 / 84.0, 44.1 / 61.1, 97.5 / 133 47.2 / 79.6, 41.2 / 61.2, 94.5 / 127 −5%, +0.2%, −4.5% flat or better
blaze long_task_max, heap_growth, requests, blank_frames, cls 207, 1.84, 52, 0, 0 204, 1.89, 52, 0, 0 flat

Bundle (brotli KB, perf/web.sh --bundle-only, deterministic): first JS 233.20 → 233.95 (+0.75, +0.3%); first CSS 18.80 → 19.59 (+0.79, +4.2%); total JS 353.3 → 364.6 (+11.3, idle-screens +10.4); all CSS 34.5 → 37.0. All inside budgets (first JS 240, first CSS 22.1) but larger than dev.

Perf verdict: RED. Beyond noise: sync.round_trip p95 +52%/+58% (≈ +0.2 ms, 10 ms budget), cold first contentful paint / first note +84 / +67 ms p50 at 4× slowdown, list_search.key p95 +14% (+1.5 ms), and a larger first CSS (+4.2%). Not bisected: sync.round_trip points at the #829/#835 room/live changes, the browser numbers at the #850–#852 style stack (bundle +11 KB). Everything else is flat or better.

Summary

Gates pass; the matrix has 10 train-only deterministic failures (Style sheet, Cover, Width, Opacity) plus vim:380 that fails on dev too and one first-run accent-labels flake that did not reproduce; perf has the regressions above. Artifacts (JSON, logs, reports) are only in these comments; VM work dirs are cleaned.

## BM14c result: RED. Do not merge. Head `b59e4a5e2155b2b71c6246b98a4e20679073cf26`, base origin/dev `41f2efa387b74175971a838f602535089d0d5fb4`. Merged `--no-ff` in order, heads verified first: #850 `af6fb861`, #851 `d94771b6`, #852 `656b61bf`, #829 `97d790f6`, #835 `3755691d` (current), #895 `3c1d8deb`. No conflicts. `CI=1`, `SLOT_TRAIN=1`, train lock with EXIT trap on both VMs, on-disk TMPDIR, `--workers=1 --retries=0`, df before each step (dev-01 ≥ 235 GiB, dev-02 ≥ 300 GiB free). First run counted, nothing retried. No fixes made. ## Gates (first run) | Gate | Command | Result | |---|---|---| | install (both VMs) | `bun install --frozen-lockfile` | pass | | fmt | `cargo fmt --all --check` | pass | | clippy | `cargo clippy --workspace --all-targets --locked -j4 -- -D warnings` | pass | | Rust | `cargo test --workspace -j4` | pass | | cargo build (both) | `cargo build -j4` | pass | | web build (both) | `cd apps/web && bun run build` | pass | | bundle budgets | `PERF_DIST=$PWD/apps/web/dist perf/web.sh --bundle-only --gate` | pass | | case paths | `node ci/check-case-paths.mjs` | pass | | evidence | `node ci/check-no-evidence.mjs` | pass | | `apps/web` `bun run check` | | **fails** at `actions.gen.ts is out of date with contracts/actions.json`. The same failure happens on origin/dev `41f2efa3` (built the same way), so it is pre-existing, not the train. | | `svelte-check` run directly (the `check` script stops before it) | `bunx svelte-check --tsconfig ./tsconfig.json` | train: 3 errors; dev: 2. The train-only error: `src/components/style/StyleSheet.svelte:142:158 Property 'noteId' is missing in type '{}' but required` (`<PageTab />`). The 2 in `e2e/review4-signout.spec.ts:225-226` are on dev too. | ## Playwright matrix (every project, one attempt each, retry 0) Chromium half on dev-01, WebKit half on dev-02, `admin-setup` on each, via the `.mts` wrapper as in BM14b. | VM / project | Passed | Failed | Skipped | |---|---:|---:|---:| | chromium / chromium-1440 | 433 | **10** | 625 | | chromium / chromium-390 | 892 | **7** | 169 | | chromium / throttled-1440 * | 3 | 0 | 2 | | chromium / throttled-390 * | 5 | 0 | 0 | | chromium / admin-chromium-1440, -390 | 12, 12 | 0 | 9, 9 | | chromium / admin-setup, flow, passkeys | 1, 1, 1 | 0 | 0 | | chromium / prefs-chromium | 7 | 0 | 0 | | chromium / look-chromium-1440, -390 | 5, 5 | 0 | 0 | | chromium / contrast | 9 | 0 | 0 | | chromium / sidebar-band-chromium | 3 | 0 | 0 | | webkit / webkit-1440 | 430 | **10** | 628 | | webkit / webkit-390 | 888 | **7** | 173 | | webkit / admin-webkit-1440, -390 | 9, 9 | 0 | 9, 9 | | webkit / admin-setup | 1 | 0 | 0 | | webkit / prefs-webkit | 3 | 0 | 4 | | webkit / look-webkit-1440, -390 | 4, 4 | 0 | 1, 1 | | webkit / sidebar-band-webkit | 3 | 0 | 0 | | **Total** | **2732 + 8\*** | **34** | | \* The throttled projects depend on the failed chromium projects, so Playwright skipped them. Their first attempts were run right after with `--no-deps` on the same train binary and dist. Known flakes `sidebar-edge.spec.ts:8`, webkit-1440 `editor.spec.ts:173`, webkit-390 `missing-links.spec.ts:245` did not occur. ## Failures (34 first-run failures = 17 cases on each browser) One comparison on origin/dev `41f2efa3` (built and run the same way, exact case, one attempt) and `--repeat-each=10 --no-deps` on the train. "pass n/10" is the train repeat. | Case | Projects failed in matrix | First error | origin/dev | Train repeat | |---|---|---|---|---| | `cover.spec.ts:124` Style sheet Cover row | all four (1440, 390, both browsers) | `[data-testid=style-cover]` Change button not found | spec does not exist on dev | 0/10 on each of the four | | `cover.spec.ts:143` "Use this picture as the cover" | all four | `style-cover` "No cover" not found | spec does not exist on dev | 0/10 on each | | `style-sheet.spec.ts:26` | all four | test timeout 30 s | spec does not exist on dev | 0/10 on each | | `style-sheet.spec.ts:49` | all four | test timeout 30 s | spec does not exist on dev | 0/10 on each | | `note-bars.spec.ts:319` Style popover tabs | 1440 chromium, webkit | timeout 30 s | spec does not exist on dev | 0/10 on each | | `note-bars.spec.ts:237` Style tabs follow height | 390 chromium, webkit | "the Text page is shorter than Page": expected < 245, got 394 | spec does not exist on dev | chromium 0/10, webkit 1/10 | | `opacity.spec.ts:122` (dev `:120`) slider to 0% | all four | timeout 30 s | **pass** on all four | 0/10 on each | | `width.spec.ts:48` | 1440 chromium, webkit | timeout 30 s | **pass** (both) | 0/10 on each | | `width.spec.ts:86` | 1440 chromium, webkit | tab "Wide" not found | **pass** (both) | 0/10 on each | | `width.spec.ts:177` | all four | tab "Narrow" not found (900px) | **pass** on all four | 0/10 on each | | `vim.spec.ts:387` (dev `:380`) `:` commands | chromium-1440 | `Saved` not visible | **fails on dev too** (reproduces on baseline) | 6/10 pass: flaky on both, not the train | | `accent-labels.spec.ts:26` | webkit-1440 | accent radios: 0 found, expected 7 | pass | 10/10 pass: first-run flake, not reproduced | So 10 distinct cases fail on every repeat only on the train (cover ×2, style-sheet ×2, note-bars ×2, opacity, width ×3). `vim:380/387` fails on dev too; `accent-labels:26` did not reproduce. Likely cause (not fixed): `StyleSheet.svelte` renders `<PageTab />` without the required `noteId` after the #850/#851/#852 merge (svelte-check error above); the missing Width, Cover and Opacity rows are on that tab. ## Perf (new gate: no regression vs origin/dev) Train `b59e4a5e` against origin/dev `41f2efa3`, release binaries built on dev-01 and relayed with matching SHA-256 on both perf VMs (head `calternal` `56787f47…`, dev `2c6a9b93…`, same `calternal-perf` `4631c52b…`). Every run under `~/bin/perf-lock`, seed 1, pairs alternating, perf VM loads quiet (0.3–5.2). **Server, perf-02** (`calternal-perf --sizes 10k,100k --runs 5 --cold-runs 0 --idle-secs 0 --storm-edits 1`, 3 alternating pairs; medians of per-run p50 / p95; spread = min–max of per-run p95 across the 3 runs): | Metric | Home | dev p50 / p95 | head p50 / p95 | p95 delta | dev spread | head spread | Verdict | |---|---|---|---|---:|---|---|---| | **sync.round_trip** (ms) | 10k | 0.34 / 0.42 | 0.49 / 0.64 | **+52%** | 0.39–0.46 | 0.63–0.65 | **regression** (ranges do not overlap) | | **sync.round_trip** (ms) | 100k | 0.35 / 0.45 | 0.57 / 0.71 | **+58%** | 0.42–0.50 | 0.68–0.81 | **regression** | | open.note (ms) | 10k | 0.91 / 1.84 | 1.28 / 2.06 | +12% | 1.46–2.22 | 1.79–2.17 | within noise (p50 +0.37 ms, watch) | | open.note (ms) | 100k | 5.64 / 12.3 | 6.04 / 10.6 | −14% | 11.2–13.4 | 8.45–13.1 | flat | | open.api_get (ms) | 10k / 100k | 2.94 / 10.6, 6.59 / 12.1 | 3.04 / 12.7, 6.16 / 13.5 | +20% / +12% | 6.3–11.6, 11.7–19.0 | 11.4–20.1, 9.6–16.5 | within noise | | home.page / page_title / page_next / page_tag (ms) | 10k | 0.88 / 1.78, 0.84 / 1.06, 0.94 / 1.27, 6.09 / 7.78 | 0.85 / 2.21, 0.68 / 0.99, 0.91 / 1.31, 5.75 / 7.92 | +24%, −7%, +3%, +2% | wide | | within noise | | home.page / page_title / page_next / page_tag (ms) | 100k | 1.07 / 2.63, 0.85 / 1.71, 0.98 / 1.60, 57.9 / 65.3 | 1.23 / 1.83, 0.92 / 1.51, 1.08 / 1.37, 61.1 / 69.1 | −30%, −11%, −15%, +6% | page_tag 61.9–68.6 | 66.8–70.9 | within noise; page_tag budget 50 ms is over on dev too | | home.first_page (ms) | 10k / 100k | 2.86 / 5.32, 4.20 / 6.81 | 2.79 / 6.11, 5.35 / 9.19 | +15% / +35% | 3.1–6.8, 4.6–12.9 | 4.4–6.6, 6.5–12.1 | within noise (single sample per restart) | | search.query (ms, prefix/typo mix of the standard harness) | 10k / 100k | 1.94 / 4.56, 2.26 / 12.3 | 1.71 / 4.38, 2.32 / 5.94 | −4% / −52% | | | flat or better | | search.first_query (ms) | 10k / 100k | 5.7 / 32.8, 8.9 / 52.0 | 5.4 / 24.8, 7.3 / 31.0 | −24% / −40% | | | flat or better | | sync.save_ack (ms) | 10k | 2.70 / 3.15 | 2.93 / 4.98 | +58% | 3.0–3.3 | 3.25–7.5 | one slow run; p50 +0.23 ms | | sync.save_ack (ms) | 100k | 3.54 / 5.11 | 3.85 / 5.27 | +3% | | | within noise | | sync.durable_write, journal.catch_up, startup, shutdown, mem.idle/peak | both | | | −5%…+4% (shutdown 10k +32%: 64 vs 84.5 ms, both runs bimodal 64/85 on dev too) | | | flat | **200 appends** (30 alternating pairs, perf-01, `CALTERNAL_NOTES_SAVE_MS=15`, same `append.py` as #792): total time median dev 1.308 s, head 1.328 s = **+1.50%**; per-append p95 median 11.55 → 12.10 ms (+4.8%); pair ratios 0.79–1.21. Noise control, the same script with the dev binary against a copy of itself (30 pairs): **+2.11%**. So the +1.50% is inside run-to-run noise. **Browser, perf-01** (Chromium 4× CPU slowdown, 10k Home, `perf/boot.perf.ts`, `list-search.perf.ts`, `blaze.perf.ts`; 3 alternating runs per side; each side served by its own server binary and dist; WebKit not available on perf VMs): | Metric | dev p50 / p95 | head p50 / p95 | p95 delta | per-run medians dev / head | Verdict | |---|---|---|---:|---|---| | **boot.cold_fcp** (ms) | 820 / 916 | 904 / 956 | **+4.4%** (p50 +10%) | 804–828 / 860–936 | **regression** (separated) | | **boot.cold_note** (first note painted, ms) | 1051 / 1186 | 1118 / 1204 | +1.5% (p50 +6.3%) | 1017–1073 / 1113–1198 | **regression** at p50 (separated), p95 within noise | | boot.cold_list (ms) | 1051 / 1186 | 1113 / 1199 | +1.1% (p50 +5.9%) | 1017–1073 / 1027–1180 | within noise | | boot.cold_tti / cold_cpu / cold_tbt | 1893 / 2507, 1477 / 1732, 199 / 254 | 1980 / 2496, 1632 / 1743, 187 / 230 | −0.4%, +0.6%, −9% | | within noise | | boot.warm_note / warm_list / warm_tti (ms) | 531 / 971, 357 / 707, 1496 / 1980 | 534 / 977, 355 / 717, 1480 / 1991 | +0.6%, +1.4%, +0.5% | | flat | | **list_search.key** (keystroke to paint, ms) | 7.6 / 10.6 | 7.9 / 12.1 | **+14%** (p50 +3.9%) | 7.5–7.8 / 7.8–8.2 | **small regression** (separated) | | blaze.shown / back_shown / live (ms) | 47.5 / 84.0, 44.1 / 61.1, 97.5 / 133 | 47.2 / 79.6, 41.2 / 61.2, 94.5 / 127 | −5%, +0.2%, −4.5% | | flat or better | | blaze long_task_max, heap_growth, requests, blank_frames, cls | 207, 1.84, 52, 0, 0 | 204, 1.89, 52, 0, 0 | | | flat | **Bundle** (brotli KB, `perf/web.sh --bundle-only`, deterministic): first JS 233.20 → 233.95 (+0.75, +0.3%); first CSS 18.80 → 19.59 (+0.79, **+4.2%**); total JS 353.3 → 364.6 (+11.3, idle-screens +10.4); all CSS 34.5 → 37.0. All inside budgets (first JS 240, first CSS 22.1) but larger than dev. **Perf verdict: RED.** Beyond noise: `sync.round_trip` p95 +52%/+58% (≈ +0.2 ms, 10 ms budget), cold first contentful paint / first note +84 / +67 ms p50 at 4× slowdown, `list_search.key` p95 +14% (+1.5 ms), and a larger first CSS (+4.2%). Not bisected: `sync.round_trip` points at the #829/#835 room/live changes, the browser numbers at the #850–#852 style stack (bundle +11 KB). Everything else is flat or better. ## Summary Gates pass; the matrix has 10 train-only deterministic failures (Style sheet, Cover, Width, Opacity) plus `vim:380` that fails on dev too and one first-run `accent-labels` flake that did not reproduce; perf has the regressions above. Artifacts (JSON, logs, reports) are only in these comments; VM work dirs are cleaned.
kayg commented 2026-11-03 19:59:29 +00:00
Author
Owner
Copy link

BM14c Style stack: causes and fixes (pushed to #852 port/style-3, head 07b6770a)

Cause 1 (most of the red). #851 gave PageTab a noteId prop and #852's merge of #851/#850 took the <PageTab /> call from the #851 side of StyleSheet.svelte:142 (no noteId), while #850's side had <PageTab {noteId} />. So the Cover row got no note, and cover / Width / Opacity / style-sheet / note-bars specs lost their rows and timed out. svelte-check showed it as Property 'noteId' is missing (the only train-only error). Fix: f757dc1e passes noteId.

Cause 2 (note-bars.spec.ts:237, phone, WebKit 13/20 after cause 1 was fixed). The Page tab builds itself in steps (looks, then rows + Cover, then colour; 40 ms timers). StyleSheet dropped its held room (min-height) as soon as the tab was mounted, so the capsule shrank to the looks (307 px) and grew back; the spec's still() saw the shrunken moment and compared "Text page shorter than Page" against 307 (Text is 394). Fixes: fed76ea1 keeps the held room until PageTab says its last step is in (onready), and 07b6770a updates the held-room guesses for the taller Page tab with the Cover row (measured on the stack: page 464 phone / 553 wide, sheet 516 / 605; they were 362/451 and 414/503 from before the Cover row). No assertion, threshold or timeout changed.

Results on the merged train tree + these 3 commits (train b59e4a5e plus cherry-picks, dev-01, CI=1 --workers=1 --retries=0 --repeat-each=20)

Cases: cover.spec.ts:124, :143; style-sheet.spec.ts:26, :49; note-bars.spec.ts:319 (desktop), :237 (phone); opacity.spec.ts:122; width.spec.ts:177, :48 and :86 (1440 only). Cases that do not apply to a width are skipped by their own test.skip.

project failed passed skipped (by the spec)
chromium-1440 0 180 20
chromium-390 0 140 60
webkit-1440 0 180 20
webkit-390 0 140 60

Notes on the runs: the very first parallel runs shared one test-results dir and a port, giving ENOENT/port errors (my harness, not the app); every number above is from runs with their own --output. One webkit-1440 run beside three other 20x runs had 1 failure in 180 (note-bars.spec.ts:319, tablist not up in 5 s under heavy load); the same project alone, repeated: 180/180. svelte-check on the stack: only the 2 errors in e2e/review4-signout.spec.ts:225-226 that origin/dev has too; StyleSheet.svelte:142 is gone.

Style sheet open time (kept)

note-bars-motion.spec.ts "rest to Style sheet", throttled-390 (6x CPU), --no-deps --repeat-each=20: 20/20 pass; longest main-thread task of the open: min 80, median 92, p90 128, max 172 ms (limit 250; goal median ≤ 120). Held room did not add any task.

Not done

vim.spec.ts:387 (:w says Saved): not touched. It fails on dev too and I did not find a quick deterministic fix; left as is for #895.

## BM14c Style stack: causes and fixes (pushed to #852 `port/style-3`, head `07b6770a`) **Cause 1 (most of the red).** #851 gave `PageTab` a `noteId` prop and #852's merge of #851/#850 took the `<PageTab />` call from the #851 side of `StyleSheet.svelte:142` (no `noteId`), while #850's side had `<PageTab {noteId} />`. So the Cover row got no note, and cover / Width / Opacity / style-sheet / note-bars specs lost their rows and timed out. `svelte-check` showed it as `Property 'noteId' is missing` (the only train-only error). Fix: `f757dc1e` passes `noteId`. **Cause 2 (`note-bars.spec.ts:237`, phone, WebKit 13/20 after cause 1 was fixed).** The Page tab builds itself in steps (looks, then rows + Cover, then colour; 40 ms timers). `StyleSheet` dropped its held room (`min-height`) as soon as the tab was *mounted*, so the capsule shrank to the looks (307 px) and grew back; the spec's `still()` saw the shrunken moment and compared "Text page shorter than Page" against 307 (Text is 394). Fixes: `fed76ea1` keeps the held room until PageTab says its last step is in (`onready`), and `07b6770a` updates the held-room guesses for the taller Page tab with the Cover row (measured on the stack: page 464 phone / 553 wide, sheet 516 / 605; they were 362/451 and 414/503 from before the Cover row). No assertion, threshold or timeout changed. ## Results on the merged train tree + these 3 commits (train `b59e4a5e` plus cherry-picks, dev-01, `CI=1 --workers=1 --retries=0 --repeat-each=20`) Cases: cover.spec.ts:124, :143; style-sheet.spec.ts:26, :49; note-bars.spec.ts:319 (desktop), :237 (phone); opacity.spec.ts:122; width.spec.ts:177, :48 and :86 (1440 only). Cases that do not apply to a width are skipped by their own `test.skip`. | project | failed | passed | skipped (by the spec) | |---|---|---|---| | chromium-1440 | 0 | 180 | 20 | | chromium-390 | 0 | 140 | 60 | | webkit-1440 | 0 | 180 | 20 | | webkit-390 | 0 | 140 | 60 | Notes on the runs: the very first parallel runs shared one `test-results` dir and a port, giving ENOENT/port errors (my harness, not the app); every number above is from runs with their own `--output`. One webkit-1440 run beside three other 20x runs had 1 failure in 180 (`note-bars.spec.ts:319`, tablist not up in 5 s under heavy load); the same project alone, repeated: 180/180. `svelte-check` on the stack: only the 2 errors in `e2e/review4-signout.spec.ts:225-226` that origin/dev has too; `StyleSheet.svelte:142` is gone. ## Style sheet open time (kept) `note-bars-motion.spec.ts` "rest to Style sheet", throttled-390 (6x CPU), `--no-deps --repeat-each=20`: 20/20 pass; longest main-thread task of the open: min 80, **median 92**, p90 128, max 172 ms (limit 250; goal median ≤ 120). Held room did not add any task. ## Not done `vim.spec.ts:387` (`:w` says Saved): not touched. It fails on dev too and I did not find a quick deterministic fix; left as is for #895.
kayg commented 2026-11-03 19:59:40 +00:00
Author
Owner
Copy link

Correction to cause 1: #850 passes <PageTab {noteId} />; #851's StyleSheet.svelte rewrote that call as <PageTab /> (its PageTab takes no prop); #852 made PageTab require noteId for the Cover row and kept #851's call. So the missing prop is a #851/#852 interaction, fixed on #852 by f757dc1e.

Correction to cause 1: #850 passes `<PageTab {noteId} />`; #851's `StyleSheet.svelte` rewrote that call as `<PageTab />` (its PageTab takes no prop); #852 made `PageTab` require `noteId` for the Cover row and kept #851's call. So the missing prop is a #851/#852 interaction, fixed on #852 by `f757dc1e`.
kayg commented 2026-11-04 00:41:14 +00:00
Author
Owner
Copy link

BM14c perf bisect and fixes (not merged)

Method: release binaries built on dev-01, relayed to the perf VMs with matching sha256 (dev de99cfbf…, train+fixes 68dd30a0…, same dev calternal-perf for every run). Every run under ~/bin/perf-lock, seed 1, 10k Home, alternating pairs. Medians of per-run p50 / p95, [min–max of p50]. The perf VMs were shared with other agents' long lock holds, so runs spread over several windows; dev-vs-dev spread alone is 0.20–0.37 ms on round trip.

sync.round_trip bisect (10k, perf-02, ms)

Build p50 p95 runs
dev 41f2efa3 0.302–0.327 0.41–0.43 5
train through #829 (ef284146) 0.418 (+34%) 0.547 (+33%) 5
train through #835 (efa2bf37) 0.548 (+77%) 0.766 (+87%) 4
#829 + fix (72eb0eb1) vs dev same window 0.217 vs 0.235 0.337 vs 0.344 5
#835 + owner-delivery fixes only (F3) 0.351 vs 0.327 (+7%) 0.451 vs 0.417 5
train + all fixes (TF2) vs dev, 10k 0.286 vs 0.290 0.388 vs 0.401 5

Culprits: #829 (Auth::authorize_room_delivery, called per delivered frame, does a sharing-db read even for the owner's own note) and #835 (the same check per queued frame in the room, again per written frame in live.rs, plus two more reads per room join and a clone of the caller). Both are the same cause: an owner reading their own note never depends on a grant row, but every frame paid a database hop.

Final train + fixes vs dev, both sizes (3 pairs, ms; p50 / p95)

Metric Home dev train+fixes note
sync.round_trip 10k 0.214 / 0.351 0.231 / 0.341 p95 flat, p50 +8% inside range overlap
sync.round_trip 100k 0.246 / 0.353 0.288 / 0.379 p50 +17%, p95 +7%: n=3, ranges overlap, not shown flat
sync.save_ack 10k / 100k 2.63 / 4.41, 3.31 / 5.11 3.11 / 4.17, 3.27 / 4.68 flat or better
open.note 10k 0.710 / 1.183 0.787 / 1.408 +11% / +19%, n=3
open.note 100k 4.81 / 9.49 4.47 / 10.64 flat / +12%
open.api_get 10k / 100k 3.47 / 9.94, 5.04 / 12.2 3.31 / 10.95, 5.12 / 10.9 flat
Earlier 5-pair 10k run of the same binary: round_trip 0.286 / 0.388 vs dev 0.290 / 0.401, open.note 0.817 / 1.531 vs 0.815 / 1.440 (flat). I could not get more 100k pairs through the shared lock; 100k round_trip needs a longer paired run before it can be called flat.

Browser (Chromium 4x CPU, 10k Home; boot.perf.ts, list-search.perf.ts)

Pooled, interleaved dev / #850 / #851 / #852 / #852+CSS-split runs across three windows (D n=21, #850 n=7, #851 n=5, #852 n=14, split n=7):

Metric (p50 ms) dev #850 #851 #852 #852 + CSS split
boot.cold_fcp 744 780 736 760 748
boot.cold_note 938 988 927 951 968
list_search.key (p50 / p95) 5.9 / 8.2 6.2 / 8.3 5.9 / 8.6 6.2 / 8.3 6.1 / 8.9
The train's +10% cold FCP, +6% first note and +14% search-key p95 did not reproduce: in the quietest window (8 interleaved rounds) #852 was flat to dev (FCP 744 vs 756, first note 941 vs 962, search key 6.0 / 8.1 vs 5.8 / 8.0), and dev-vs-dev spread is 704–832 ms FCP. #850 shows the largest step (+2.6…4.8%, inside range overlap). I found no per-keystroke work added to the search path.
Bundle (brotli KB, bundle-size.mjs): first CSS dev 18.8 → train 19.6 → with split 19.2; first JS 233.2 → 233.9 → 233.6. The remaining +0.4 KB CSS is the rest bar's own rules, needed by the first note screen.

Fix branches pushed to the owning PRs: #829 72eb0eb1, #835 46be23e3, #850 6eb81f62. Train order check: #852 state + #829 + #835 + #895 merges cleanly (no conflicts) with all fixes; tree = train head + 5 changed files.

## BM14c perf bisect and fixes (not merged) Method: release binaries built on dev-01, relayed to the perf VMs with matching sha256 (dev `de99cfbf…`, train+fixes `68dd30a0…`, same dev `calternal-perf` for every run). Every run under `~/bin/perf-lock`, seed 1, 10k Home, alternating pairs. Medians of per-run p50 / p95, [min–max of p50]. The perf VMs were shared with other agents' long lock holds, so runs spread over several windows; dev-vs-dev spread alone is 0.20–0.37 ms on round trip. ### sync.round_trip bisect (10k, perf-02, ms) | Build | p50 | p95 | runs | |---|---|---|---| | dev `41f2efa3` | 0.302–0.327 | 0.41–0.43 | 5 | | train through #829 (`ef284146`) | 0.418 (+34%) | 0.547 (+33%) | 5 | | train through #835 (`efa2bf37`) | 0.548 (+77%) | 0.766 (+87%) | 4 | | #829 + fix (`72eb0eb1`) vs dev same window | 0.217 vs 0.235 | 0.337 vs 0.344 | 5 | | #835 + owner-delivery fixes only (F3) | 0.351 vs 0.327 (+7%) | 0.451 vs 0.417 | 5 | | **train + all fixes (TF2) vs dev, 10k** | **0.286 vs 0.290** | **0.388 vs 0.401** | 5 | Culprits: **#829** (`Auth::authorize_room_delivery`, called per delivered frame, does a sharing-db read even for the owner's own note) and **#835** (the same check per queued frame in the room, again per written frame in `live.rs`, plus two more reads per room join and a clone of the caller). Both are the same cause: an owner reading their own note never depends on a grant row, but every frame paid a database hop. ### Final train + fixes vs dev, both sizes (3 pairs, ms; p50 / p95) | Metric | Home | dev | train+fixes | note | |---|---|---|---|---| | sync.round_trip | 10k | 0.214 / 0.351 | 0.231 / 0.341 | p95 flat, p50 +8% inside range overlap | | sync.round_trip | 100k | 0.246 / 0.353 | 0.288 / 0.379 | **p50 +17%, p95 +7%: n=3, ranges overlap, not shown flat** | | sync.save_ack | 10k / 100k | 2.63 / 4.41, 3.31 / 5.11 | 3.11 / 4.17, 3.27 / 4.68 | flat or better | | open.note | 10k | 0.710 / 1.183 | 0.787 / 1.408 | +11% / +19%, n=3 | | open.note | 100k | 4.81 / 9.49 | 4.47 / 10.64 | flat / +12% | | open.api_get | 10k / 100k | 3.47 / 9.94, 5.04 / 12.2 | 3.31 / 10.95, 5.12 / 10.9 | flat | Earlier 5-pair 10k run of the same binary: round_trip 0.286 / 0.388 vs dev 0.290 / 0.401, open.note 0.817 / 1.531 vs 0.815 / 1.440 (flat). I could not get more 100k pairs through the shared lock; 100k round_trip needs a longer paired run before it can be called flat. ### Browser (Chromium 4x CPU, 10k Home; `boot.perf.ts`, `list-search.perf.ts`) Pooled, interleaved dev / #850 / #851 / #852 / #852+CSS-split runs across three windows (D n=21, #850 n=7, #851 n=5, #852 n=14, split n=7): | Metric (p50 ms) | dev | #850 | #851 | #852 | #852 + CSS split | |---|---|---|---|---|---| | boot.cold_fcp | 744 | 780 | 736 | 760 | 748 | | boot.cold_note | 938 | 988 | 927 | 951 | 968 | | list_search.key (p50 / p95) | 5.9 / 8.2 | 6.2 / 8.3 | 5.9 / 8.6 | 6.2 / 8.3 | 6.1 / 8.9 | The train's +10% cold FCP, +6% first note and +14% search-key p95 did **not** reproduce: in the quietest window (8 interleaved rounds) #852 was flat to dev (FCP 744 vs 756, first note 941 vs 962, search key 6.0 / 8.1 vs 5.8 / 8.0), and dev-vs-dev spread is 704–832 ms FCP. #850 shows the largest step (+2.6…4.8%, inside range overlap). I found no per-keystroke work added to the search path. Bundle (brotli KB, `bundle-size.mjs`): first CSS dev 18.8 → train 19.6 → with split 19.2; first JS 233.2 → 233.9 → 233.6. The remaining +0.4 KB CSS is the rest bar's own rules, needed by the first note screen. Fix branches pushed to the owning PRs: #829 `72eb0eb1`, #835 `46be23e3`, #850 `6eb81f62`. Train order check: #852 state + #829 + #835 + #895 merges cleanly (no conflicts) with all fixes; tree = train head + 5 changed files.
kayg commented 2026-11-07 17:53:13 +00:00
Author
Owner
Copy link

Train superseded by #914 (BM14e), landed in dev via 12103f1c97. Closing the obsolete BM14c train.

Train superseded by #914 (BM14e), landed in dev via 12103f1c973f1e810dd4a9a3bff5a705514d70b0. Closing the obsolete BM14c train.
kayg closed this pull request 2026-11-07 17:53:14 +00:00
Some checks failed
CI / server (pull_request) Has been cancelled
Details
CI / rust (pull_request) Has been cancelled
Details
CI / web (pull_request) Has been cancelled
Details
Markdown extraction contract / markdown-native-contract (pull_request) Has been cancelled
Details
Markdown extraction contract / markdown-editor-contract (pull_request) Has been cancelled
Details

Pull request closed

Please reopen this pull request to perform a merge.
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels   
backend
Rust server

  
design
Mockups, UI and TASTE

  
docs
ADRs, CONTEXT, TASTE

  
idea
Parked idea, never worked until promoted

  
infra
Deploy, CI, VMs

  
perf
Performance and budgets

  
reliability
Data integrity, corruption traps

  
web
Web client

No labels
backend
design
docs
idea
infra
perf
reliability
web
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!897
No description provided.