chore(memory): V1-2.2 three MEDIUM evidence, lot B patch, contract, TODO, journal
This commit is contained in:
1 parent
9e49ed1b4d
commit
f1bb19b30b
9 files changed
+1652
No files matched your search
@@ -0,0 +1,82 @@
|
||||
# V1-2.2 three MEDIUM: evidence (2026-10-30)
|
||||
|
||||
Owner brief « V1-2.2 : arbitrages propriétaire sur les trois MEDIUM » on `958228a`. Contract
|
||||
`.claude/tasks/contracts/2026-10-30-v1-2-2-three-medium.md`. Lot A (defect 3, frame) committed `0c87c7f`, `faa53c7`,
|
||||
`587cbb2`. Lot B (defects 1 and 2, isolation and hitbox) NOT committed: it changes what frozen R3 references compare;
|
||||
patch `2026-10-30-v1-2-2-lot-b-isolation.patch` (against `587cbb2`), docs RUST.md §12.11 (lot A) and §12.12 (lot B, in
|
||||
the patch).
|
||||
|
||||
## 1. Audit
|
||||
|
||||
- Reaching the layer before: the root's presentation attributes the reader keeps. Inherited: `stroke*`, `fill*`,
|
||||
`visibility`, `color`, `clip-rule`, `color-interpolation-filters` (the halo filter sat in the drawing's first
|
||||
`<defs>`), font properties. Whole subtree, no child can undo: `opacity`, `display`, `systemLanguage`. Already removed
|
||||
on the root by the reader: `class`, `style`, `transform`, `filter`, `mask`, `clip-path`; no `pointer-events`,
|
||||
`cursor` or `<style>` survives anywhere; root `visibility` refused (`root_visibility`).
|
||||
- Markers: point hitbox `fill` transparent, `pointer-events` all, no stroke (inherits it, and under `all` the stroke
|
||||
area is hit even unpainted); visible twin `pointer-events` none; zones and trigger declare `stroke-width` 2, not
|
||||
`stroke-opacity`, `stroke-dasharray` (zones), `stroke-linejoin`, `fill-rule`.
|
||||
- Frame: canvas = view box, else drawing `viewBox` verbatim, else `0 0 100 100`; PDF = view box, else the drawing's;
|
||||
root `preserveAspectRatio` kept (a `slice` crops where the PDF raster aspect is clamped to [0.2, 2] or the canvas box
|
||||
differs).
|
||||
- Equine: all 28 annotations in frame, minimum margin 34 units; each point centre hit by its own marker in Chromium.
|
||||
|
||||
## 2. Lot A, frame (committed)
|
||||
|
||||
- `sheet_frame` 12 tests: on `958228a` 9 fail, 3 accept-only pass; green after. Reported drawing (horse `viewBox` +10 000,
|
||||
content translated back, read clean) refused with its 19 main-view annotations, failures exactly
|
||||
`["verification.validation.out_of_frame"]`, document text unchanged.
|
||||
- `path.rs` unit tests: 500 random cubics, 500 random arcs against 20 001 samples; grammars.
|
||||
- Mutations F1-F16 (+F9b): 17/17 killed after pinning the failure list (F16 survived the first run).
|
||||
- Oracles: none changed. `check:all` green on the lot A tree (cargo 2 145 passed summed over the feature passes, ts-rs
|
||||
types unchanged, core-wasm 65, supabase 128, app 566, lint, build, e2e:ui 26, app image, worker tests and image;
|
||||
e2e:shell not run).
|
||||
- Module: 560 377 → 572 993 B gzip -9; Node's gzip level 9 582 919 B (budget 600 000, margin 17 081 B).
|
||||
|
||||
## 3. Lot B, isolation and hitbox (proposal)
|
||||
|
||||
- `svg_marker_isolation` 5 tests, all red on `958228a`, green with lot B.
|
||||
- Pixels (Playwright Chromium 156, Firefox 157, WebKit 27.2; PDF = sized SVG → Image → white canvas as
|
||||
`browserRasterizer`, canvas = inline SVG 1000 × 786; equine main view, `hock`, `hip`, `lumbar`, `head-trigger`
|
||||
selected, halo on). Plain equine: 0 px difference in both forms, three engines. Markers alone (drawing removed) vs the
|
||||
plain markers, old / new, Chromium PDF: `stroke-width` 132 851 / 0; `stroke-opacity` 41 393 / 0; `dasharray`
|
||||
22 528 / 0; `color-interpolation-filters` (halo) 66 271 / 0; `linejoin` 19 / 0; `opacity` markers absent / 0;
|
||||
`display`, `fill`, `fill-opacity`, `fill-rule` 0 / 0. Same pattern in Firefox and WebKit, canvas and PDF; new = 0
|
||||
in every case. Side effect: root `opacity="0"` / `display="none"` now hide the drawing alone, the same in every engine.
|
||||
- Real PDF exports (demo.fr.logos, v2full.fr.logos, v2unresolved.fr.nologos; every view rasterised by Chromium):
|
||||
byte-identical to `958228a`; 15 pages identical at 100 dpi.
|
||||
- Clicks (`apps/app/e2e/ui/canvas-clicks.spec.ts`, harness `e2e/harness/canvas.{html,tsx}`; projects chromium-desktop,
|
||||
chromium-375 touch, firefox-desktop, webkit-desktop, webkit-375 touch): lot B 67 passed, 5 fixme; `958228a` 45 failed
|
||||
(Chromium and WebKit every stroke case; Firefox 3). Full `e2e:ui` on lot B: 93 passed, 5 skipped.
|
||||
- Mutations I1-I7 killed; K1, K7, K8, K10, V1, V6, V7 rerun killed.
|
||||
- References compared differently (not edited): R3 `highlight.json` (36 cases) and R3 canvas goldens
|
||||
(`canvas.{main,head}.json`, `.layers.json`): equal under the declared transforms `with_isolated_layer`,
|
||||
`withIsolatedLayer`, `withDeclaredHitboxes` (included in the patch, uncommitted). R4.1 `pdf/r3/real/manifest.json`
|
||||
call 1 `svgSha256` (`a9a3155c…` → `45fd4aaf…`): a byte digest, the core-wasm test fails until the owner decides.
|
||||
Unchanged: R3 `layer.json`, `sanitiser.json`, R4 PDF digests, step6 goldens, N-1 copies, seed review drawings
|
||||
(R3 layer kept for them), equine drawing (sha256 `0ea5202f…`) and fingerprint.
|
||||
- Module with lot B: 573 088 B gzip -9.
|
||||
|
||||
## 4. Pre-existing finding: junction points
|
||||
|
||||
Zones are painted and hit after the points. Share of each visible disc recorded by its own marker (plain equine,
|
||||
identical in the three engines): `c7-t1` 51 %, `l6-s1` 56 %, `t18-l1` 72 %, fore fetlock 92 %. In Firefox and WebKit
|
||||
the centre of `c7-t1` records « thoracic ». Prototype orders (DOM reordering, not built): points above zones → discs
|
||||
100 % (fore fetlock 92 %), zones lose 2 to 10 % to the point hitboxes; hitboxes, then zones, then capturing twins
|
||||
(« visible wins ») → discs 99-100 %, zones lose only what the discs cover (3 %). Owner decision.
|
||||
|
||||
## 5. Independent reviews (fresh, Opus: the pinned Fable model was out of credits)
|
||||
|
||||
- Security: PASS, semgrep 163 rules 0 findings. MEDIUM scale bypass, reproduced: `viewBox` × 1 000 with the content
|
||||
scaled back keeps the horse identical, every marker sub-pixel, verdict green (defect 3 by another route; owner
|
||||
decision, not built). MEDIUM offset and precision: slack grew with the offset (fixed `9e49ed1`, test + mutant F17);
|
||||
absolute 1e-12 epsilons in the curve extremes and f64 reading where browsers use f32 (needs coordinate range bounds,
|
||||
owner decision). LOW NaN dropped by `min`/`max` (fixed `9e49ed1`: non-finite bounds unreadable). LOW test lint allows
|
||||
(project convention).
|
||||
- Verifier: ECARTS(2). C8 unverifiable inside its brief (docker, mutations); FLOOR `test.skip` on the tap test, not
|
||||
authorised (removed: the desktop projects already leave `@phone` out). C1-C7, C9 met; declared transforms judged
|
||||
faithful (they only rewrite the frozen side, strict comparison kept). F16 anchor broken by rustfmt (fixed). Proof
|
||||
scripts were not in the repo: now `2026-10-30-v1-2-2-three-medium-tools/` (pixel probe, real PDF export, junction
|
||||
probe). Residual: a root `systemLanguage` now sits on the drawing's group (drawing only, markers unaffected).
|
||||
- After the fixes: mutations 32/32 killed; module 573 134 B gzip -9, Node 583 018 B; `e2e:ui` 93 passed, 5 fixme;
|
||||
app 566. `check:all` last ran in full on `faa53c7` (lot A before `9e49ed1`), not rerun after `9e49ed1`.
|
||||
Reference in new issue
Block a user