docs(docs): U2 final gate figures on 3d6aacad, review minors

This commit is contained in:
bchanot committed 2026-11-19 23:38:15 +01:00
1 parent 3d6aacad99
commit 481335f99a
4 files changed
+34 -32

No files matched your search

+2 -2
View File
@@ -50,7 +50,7 @@ Implemented (contract `.claude/tasks/contracts/2026-11-06-v1-2-6-ux-u1-1123.md`)
## U2 · Points and polygons
Implemented (plan `.claude/tasks/plans/2026-11-17-v1-2-6-u2.md`, rev 4 then rev 5 (owner decisions D1 to D5, 2026-11-19); boards `docs/design/v1-2-6-u2/`, A1 to A15 and the three adjustments). Rule numbers (R-n) are the plan's. Gates (plan Task 16) on the final head, 2026-11-19 (code at da8642dd, this page's figures added after): `pnpm check:all` exit 0 (app 2 140, core-wasm 67, supabase 182 tests, e2e:ui 253); the shell suite without `editor-equine.spec.ts` 330 passed, 17 failed, 5 not run, all in four files outside the editor (figures of the rev 4 code; figures on the final code to be added by the controller). These environment failures are named with their evidence and are not resolved; none is presented as formally independent of U2. Owner decision D5 (2026-11-19): they do not block the manual trial. `admin` (6): timeouts that the 56c75b1 baseline log also shows (sandbox data); no further attribution. `catalogue-integrity` (6): not attributed to U2. The pre-U2 commit 56c75b1 fails the same 6 of 6 tests, on the same assertion, with the same timings, on the same data (2026-11-19, one worker, read-only counts, nothing deleted); the published versions rose from 426 to about 690 in a day of runs, and the failing set grew from 4 to 6 in both trees, so volume is the inferred cause (the card list loads every card before the search narrows it past the 10 s wait). Volume causation is inferred from that trend and the identical baseline, not isolated: a low-volume run needs a reset, which D4 forbids. `catalogue` (3) and `admin-sheets-list` (2): « Target crashed » under the parallel run; both pass 16/16 with one worker (`--workers=1`, 2026-11-19); the logs show nothing more. `pnpm db:sandbox test` 105 passed, 10 skipped, and `test-db` Files=21 Tests=1958 FAIL on `014_animals_relations` (2) and `021_sheet_verification_jobs` (23), identical to before U2, which changes no SQL (run at 0aa0f6bc). The sandbox is never reset or rebuilt on this work (D4). Both opt-in measures rerun (figures in the Performance row). The owner trial is pending. The suites run on the sandbox only (`pnpm db:sandbox`).
Implemented (plan `.claude/tasks/plans/2026-11-17-v1-2-6-u2.md`, rev 4 then rev 5 (owner decisions D1 to D5, 2026-11-19); boards `docs/design/v1-2-6-u2/`, A1 to A15 and the three adjustments). Rule numbers (R-n) are the plan's. Gates (plan Task 16) on the final code, 3d6aacad, 2026-11-19: `pnpm check:all` exit 0 (app 2 158, core-wasm 67, supabase 182 tests, e2e:ui 253); the shell suite without `editor-equine.spec.ts`, on two workers (`--workers=2`; a run at the default count was stopped for lack of memory on the machine), 347 passed, 13 failed, 234 skipped, all in three files outside the editor (on the rev 4 code, da8642dd: 330 passed, 17 failed, 5 not run, in four files). These environment failures are named with their evidence and are not resolved; none is presented as formally independent of U2. Owner decision D5 (2026-11-19): they do not block the manual trial. `admin` (6): timeouts that the 56c75b1 baseline log also shows (sandbox data); no further attribution. `catalogue-integrity` (6): not attributed to U2. The pre-U2 commit 56c75b1 fails the same 6 of 6 tests, on the same assertion, with the same timings, on the same data (2026-11-19, one worker, read-only counts, nothing deleted); the published versions rose from 426 to about 690 in a day of runs, and the failing set grew from 4 to 6 in both trees, so volume is the inferred cause (the card list loads every card before the search narrows it past the 10 s wait). Volume causation is inferred from that trend and the identical baseline, not isolated: a low-volume run needs a reset, which D4 forbids. `catalogue` (3) and `admin-sheets-list` (2): « Target crashed » under the parallel run; both pass 16/16 with one worker (`--workers=1`, 2026-11-19); the logs show nothing more; neither crashed in the two-worker run on 3d6aacad. `animals` (1, new on 3d6aacad): axe reports a colour contrast on the two buttons of the archive confirmation, at 375 px; intermittent, 1 failure in 12 runs of that journey (`--repeat-each=3`, one worker). A centred dialog renders as before U2 (the `Dialog.tsx` change only places an anchored one); the inferred cause is axe measuring during the entrance fade, which the helper's wait on Web Animations does not see. Inferred from the code, not reproduced on 56c75b1, so not attributed. `pnpm db:sandbox test` 105 passed, 10 skipped, and `test-db` Files=21 Tests=1958 FAIL on `014_animals_relations` (2) and `021_sheet_verification_jobs` (23), identical to before U2, which changes no SQL (run at 0aa0f6bc). The sandbox is never reset or rebuilt on this work (D4). Both opt-in measures rerun (figures in the Performance row). The owner trial is pending. The suites run on the sandbox only (`pnpm db:sandbox`).
| Kind | Tests |
|---|---|
@@ -65,7 +65,7 @@ Implemented (plan `.claude/tasks/plans/2026-11-17-v1-2-6-u2.md`, rev 4 then rev
| Keyboard | create a zone, add and remove a vertex, move the zone, validate, without a pointer; « + » inserts and Backspace removes on a focused vertex, Backspace after an undo acts on the same focused handle; Ctrl+Z on a focused handle after arrow moves only cancels the preview and a later Enter commits nothing stale; after undoing « + », « − » never removes the wrong vertex; « Désélectionner » commits a pending series |
| Paths (T16) | a path annotation gets no handle, no « + » or « − », no zone move in either mode, no keyboard move; its `d` is byte-identical after every gesture on its neighbours, a save and a reload; « Remplacer par un polygone » stays the only change: explicit, validated by the core, the path kept while tracing, on Escape and on a refusal; one undo entry restoring the exact `d`, redo restoring the polygon (owner, BDR-101) |
| Accessibility | axe on selection, tracing, move states and the dialog in read and edit modes; handle names (« Point {nom} », « Sommet {n} de {nom} », « Ajouter un sommet après le sommet {n} de {nom} », « Retirer le sommet {n} de {nom} », the handle group « Sommets de {nom} »); live-region message after a move; WCAG 2.5.7: the move mode with four arrow buttons is the single-pointer path without dragging (A11) |
| Performance | commit latency under the 500 ms budget for drag, insert and remove, a measure run by hand at the gates (`e2e/support/measure-commit.ts`, `editor-measure-commit.spec.ts`, `MEASURE_COMMIT=1`) on a run-seeded equine-sized draft, never on the historical copy nor through `editor-equine.spec.ts`. Measured 2026-11-18 on a production build under `vite preview`, Chromium mouse at 1 280 px, CPU ×4, from the release to the painted frame showing the new geometry with the undo entry; on the final code (da8642dd, 2026-11-19): median / max drag 182.5 / 199.9 ms, « + » 176.3 / 183.1 ms, « − » 173.8 / 177.1 ms (three runs each; 15b23f8 and earlier figures in plan Task 10); the drag smoothness is a measure, not an assertion (`e2e/support/measure-drag.ts`, `editor-measure-drag.spec.ts`, run by hand). Measured 2026-11-18 on a production build under `vite preview`, Chromium touch at 1 180 px, CPU ×4: frame work p95 3.9 to 4.5 ms with the loupe (technique A, final code 15b23f8; 4.2 to 4.7 ms before), 4.0 to 4.4 ms outline only (B), 3.6 to 4.2 ms without a loupe; technique A kept. Rerun on da8642dd (2026-11-19, same setup): touch with the loupe, handler to frame median 3.0 ms, p95 4.3 ms (event to frame median 27.8 ms, p95 30.3 ms); mouse, handler to frame p95 3.1 ms. The figures are in plan Task 8; the owner's tablet trial is the real check |
| Performance | commit latency under the 500 ms budget for drag, insert and remove, a measure run by hand at the gates (`e2e/support/measure-commit.ts`, `editor-measure-commit.spec.ts`, `MEASURE_COMMIT=1`) on a run-seeded equine-sized draft, never on the historical copy nor through `editor-equine.spec.ts`. Measured 2026-11-18 on a production build under `vite preview`, Chromium mouse at 1 280 px, CPU ×4, from the release to the painted frame showing the new geometry with the undo entry; on the final code (3d6aacad, 2026-11-19): median / max drag 173.2 / 199.2 ms, « + » 184.1 / 185.5 ms, « − » 168.7 / 185.6 ms (three runs each; da8642dd: 182.5 / 199.9, 176.3 / 183.1, 173.8 / 177.1 ms; 15b23f8 and earlier figures in plan Task 10); the drag smoothness is a measure, not an assertion (`e2e/support/measure-drag.ts`, `editor-measure-drag.spec.ts`, run by hand). Measured 2026-11-18 on a production build under `vite preview`, Chromium touch at 1 180 px, CPU ×4: frame work p95 3.9 to 4.5 ms with the loupe (technique A, final code 15b23f8; 4.2 to 4.7 ms before), 4.0 to 4.4 ms outline only (B), 3.6 to 4.2 ms without a loupe; technique A kept. Rerun on 3d6aacad (2026-11-19, same setup): touch with the loupe (55 of 57 samples), handler to frame median 3.0 ms, p95 4.3 ms (event to frame median 26.0 ms, p95 36.2 ms); mouse, handler to frame p95 4.0 ms (da8642dd: 3.0 / 4.3 ms, event 27.8 / 30.3 ms, mouse 3.1 ms). The figures are in plan Task 8; the owner's tablet trial is the real check |
## U3 · Category panel and own colour