Files
faunosteo/docs/development/V1-2-6_U1_CORRECTIONS.md
T

954 lines
91 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# V1-2.6 U1 corrections after the owner's first trial
Status: PROPOSAL, 2026-11-09. Owner brief « V1-2.6 UX — Arbitrages du premier essai propriétaire » (2026-11-09),
answering the trial diagnosis of the same day. Design, documentation and plan only: nothing here authorises code.
Lot R0 (boards) is delivered in `docs/design/v1-2-6-r0/` and awaits the owner's graphic validation; R1 to R4 start
only on their own explicit authorisation. Contract: `.claude/tasks/contracts/2026-11-09-v1-2-6-trial-arbitrations.md`.
Companion: `TRIAL_STACK.md` (the dedicated manual trial stack, request §5).
## 1. Owner decisions (2026-11-09)
| # | Decision |
|---|---|
| D1 | A fifth main tab « Références » between Constats and Vérification becomes the main access to the reference library. The header « Références » button is removed. |
| D2 | A contextual panel stays, opened from a finding's detail, filtered on its citations, with « Tout afficher » leading to the tab. About 34 rem wide within the available space, a lateral slide of 150 to 200 ms, `prefers-reduced-motion` honoured. |
| D3 | The tab shows collapsible tiles: closed = short citation, type, citation count; open = full information, DOI, URL, actions. Instant search with a clear button. Sort by document order (default), A to Z, citation count or type. Sorting never changes the document; reorder commands only while document order is shown. |
| D4 | Kept: unsaved changes, focus, « Aller à », permissions, read-only mode, accessibility tests. Below 640 px, Références goes in the « Plus » menu. Boards validated before any UI code. |
| D5 | History reads « Défaire » / « Refaire » instead of « Annuler » / « Rétablir », with explicit accessible names, the keyboard shortcuts kept and the distinction from dialog « Annuler » buttons kept. |
| D6 | A « Désélectionner » button near « Supprimer ». Escape deselects when no dialog, priority panel, menu or tracing in progress consumes it. Deselecting is interface state: no document change, no history entry. A click on an empty part of the drawing is left to U2 (superseded by D6b). |
| D6b | Owner, 2026-11-12, after the U1 manual test on the trial stack: a click or a tap on the drawing anywhere but on an annotation deselects (a bone or any non-interactive SVG shape, an empty part of the drawing); a click on an annotation still selects it; a drag, a pan, a zoom or a touch navigation gesture never deselects; with nothing selected a click changes nothing; same behaviour with a mouse and on a touch screen. Joins R1, beside « Désélectionner » and Escape. |
| D7 | Sheet list: instant search with clear, sort by last modification (default) or name, the last-modification date visible, better Administration home cards. No section built on UUIDs, names, slugs or test conventions; generic for every species; no artificial test filter. |
| D8 | A dedicated, persistent manual trial stack, separate from the automated sandbox: technical plan only, nothing created, reset or deleted now (`TRIAL_STACK.md`). |
| D9 | Three separate test anomalies tracked apart from U1 and U2: A fictitious fingerprints, B working copy never archived by `sheets-admin.sandbox.ts`, C integrity failure invisible in catalogue cards (section 5). |
| D10 | Lots R0 (boards), R1 (history labels, deselect), R2 (References tab), R3 (sheet list), R4 (catalogue integrity warning, independent fix). Any SQL schema or Rust core change needs its own checkpoint. |
## 2. What these decisions supersede
| Earlier validated item | Where | New state |
|---|---|---|
| U1 AC1: « no … « Références » … tab remains » | U1 contract `.claude/tasks/contracts/2026-11-06-v1-2-6-ux-u1-1123.md:373-375` | Five tabs (D1). |
| U1 AC3: panel opened from the header button and from a finding; focus back to the opener, else the header button | same contract :381-389 | Opened from a finding only; focus back to the finding's « n référence(s) » button. Scrim, `inert` tab panels, Escape / ✕ / scrim close and read-only behaviour unchanged. |
| U1 AC4 and T15 (history part): visible « Annuler » / « Rétablir », name « Annuler la dernière modification » | contract :390-397, `V1-2-6_UX_DECISIONS.md:309-322` | « Défaire » / « Refaire » (D5). T15's form wording (« Abandonner », « Abandonner les modifications ») unchanged. |
| U1 AC10: no horizontal scroll with the references panel open | contract :420-421 | Kept, re-checked with the 34 rem panel. |
| U1 AC12, AC13: « Aller à » a reference opens the side panel | contract :425-431 | « Aller à » a reference opens the Références tab, the tile open and focused. |
| U1 AC17: panel tests « open from header and finding »; keyboard order with the panel | contract :439-446 | Tests rewritten for the tab and the finding-only panel; the panel keeps its place in the keyboard order when open. |
| Checkpoint choice 5: Général and Constats under « Plus » below 640 px | `V1-2-6_UX_U1_CHECKPOINT.md:24` | Général, Constats and Références under « Plus ». |
| Checkpoint choice 10: search on citation, DOI and URL; ↑/↓ only on the unfiltered list | same file :29 | Carried to the tab: same search fields; reorder only in document order and without an active search. |
| Checkpoint choice 15: « Références » announced as a panel | same file :34 | Announced as a tab, like the four others. |
| Board 11 (references side panel as the main access) | `docs/design/v1-2-6/11-constats-references-desktop.png` | Replaced by boards 23 to 27 once validated. |
| BDR-099 item 5 (collapsible references rejected) | memory | Not contradicted: it concerned the catalogue page on a phone (`docs/design/README.md`, accepted deviation 2), not the editor. |
| U2b F5 (filter panel in the references panel's slot) | `V1-2-6_UX_FINDINGS_NAV.md:104-126` | Unaffected: the panel slot stays, used by the contextual references panel now and by the filter panel in U2b. |
## 3. Lots
Every lot: one commit or more per coherent change, pushed; gates in this order: `pnpm typecheck`, `pnpm lint`,
`pnpm test:run`, `pnpm build`, `pnpm e2e:ui`, then the shell suite on the sandbox as it stands, with its exact result
(the four `admin.spec.ts:717` failures of the taken working-copy slot, anomaly B, are fixed: row B of section 5). Every
removed or rewritten assertion is mapped to its replacement in the commit message (R-24). Design gates: the boards of
R0 as validated, `impeccable detect` on the changed components, a fresh design review before the lot closes.
### R0 — Boards and owner validation (this delivery)
Boards 23 to 30 in `docs/design/v1-2-6-r0/` (sources, data generated from the equine draft, `REVUE.md`). Exit: the
owner validates each board, or asks for corrections; the validated boards then join `docs/design/README.md` as the
reference of R1 to R4.
### R1 — History labels and deselection (D5, D6, D6b)
Delivered 2026-11-12 on `feature/v1-2-anatomical-sheets`: `7dda61d` (D5 labels), `bbb1668` (« Désélectionner » and
Escape), `bbdfd27` + `a1ee194` (D6b), `ab823f4` + `2327d61` + `0849e54` (browser specs), `0d1edac` (design review fix:
44 px touch target below xl, 6 px icon gap), review fixes `132756e`, `6ed90ef`, `9807ed2`. Final gates on `a180b9e`
(the source at HEAD differs only by the comment of `6ef5223`): typecheck, lint, build ok; unit tests app 1381, supabase
180, core-wasm 66; `e2e:ui` 227 passed; `e2e:shell` full suite on the sandbox 197 passed, 4 failed, 99 skipped.
Status (owner, 2026-11-12): accepted technically on the report; the owner's manual validation on db:trial is still to
do. Decision record BDR-113 (R1 editor only), learnings LRN-128 to LRN-130. The 4 failures are the `admin.spec.ts:717` timeouts (anomaly B, the taken working-copy slot), the same as in
the 188 / 90 / 4 baseline; every new R1 test passed.
Follow-up `927537b` (owner brief « V1-2.6 — R1 — Dernier retour utilisateur », 2026-11-12): the « Sélection : … » tag
is screen-reader-only (`sr-only`, not a live region) in the editor and the read-only twin, so the drawing no longer
moves when the selection changes; board `04` (and board `26` of `v1-2-6-r0/`) deviation recorded in
`docs/design/README.md`; one unit test and two layout-stability specs in `editor-deselect.spec.ts` (desktop,
read-only twin at 375 px). Gates on that commit: typecheck, lint, `pnpm test:run` (app 120 files, 1382 tests) green;
`e2e:shell` on `db:sandbox`, six editor specs (deselect, equine, access, annotations, geometry, journey): 37 passed,
37 skipped (other viewport), 0 failed.
Final status (owner, 2026-11-12, brief « Validation finale R1, correctif E2E et GO R2 »): R1 VALIDATED FUNCTIONALLY.
The ten manual R1 checks on db:trial passed, and the owner confirmed in the browser that the drawing no longer moves on
selection and deselection; `927537b` and `00936c8` accepted. Limits kept: the ToolNotes hint above the canvas still
moves the drawing one line when it is cleared by a selection, the selection itself is not announced to screen readers,
iOS Safari is not validated, and the graphic gaps listed under « Behaviour and limits » and the TODO R1 follow-ups.
The criteria below are the contract as written.
- Labels: `editor.history.*` keys in `apps/app/src/i18n/fr.ts` and `en.ts` (FR « Défaire » / « Refaire », names
« Défaire la dernière modification » and « Refaire la dernière modification défaite »; EN « Undo » / « Redo »
unchanged); `HistoryButtons.tsx` keeps its structure. Shortcuts unchanged: Ctrl/⌘+Z undo, Ctrl/⌘+Maj+Z redo
(`apps/app/src/editor/useUndoShortcuts.ts:19-23`). Every prose string that names the history « Annuler » is
reworded; dialog « Annuler » buttons untouched.
- « Désélectionner »: a secondary button beside « Supprimer » (`PropertiesPanel.tsx:257-284`, removed at U2: the strip's `SelectionCommands.tsx` holds it), shown while an
annotation is selected, in read-only modes too (selection is interface state); it dispatches the existing
« no selection » path that « Toutes les annotations » uses (`PropertiesPanel.tsx:93-108`, removed at U2).
- Escape: today a tracing consumes it when the canvas box has focus (`usePlacement.ts:82`), the references panel
when not already handled (`ReferencesPanel.tsx:234`), dialogs natively (`ModalDialog.tsx:111`). One handler on the
workspace, at the bubbling phase, deselects only when `event.defaultPrevented` is false, no dialog is open, no menu
is open and no tracing runs; each consumer marks the key handled. The « Plus » menu's own Escape path is checked
first (it has no explicit handler today).
- Click on the drawing (D6b). Today, with the select tool, `clickHandler` in
`apps/app/src/editor/annotations/useCanvasGestures.ts:57-68` calls `onSelect` only when the target has a
`[data-anno-id]` ancestor (`annotationIdOf`, `useCanvasBuild.ts:91-94`) and ignores every other click: that is the
defect. Change: the other branch calls a new `onDeselect` input; `EditableAnnotations.tsx` (`CanvasStage`) wires
it to the same deselect path as the button, and only when an annotation is selected (nothing selected: no
dispatch, so a pending keyboard creation, its typed fields and the « Filtrer par nom » term stay). The read-only
twin (`AnnotationsTab.tsx:187`) draws through `CatalogueCanvas`, whose click listener
(`CatalogueCanvas.tsx:99-104`) gets an optional `onEmpty` with the same rule; the catalogue reader page
(`SheetRoute.tsx:150`) passes none and is unchanged.
- Interactive means any `[data-anno-id]` element of the composed view: a point (hitbox and disc), a zone (filled,
`pointer-events` all, `docs/ANNOTATION-RENDERING.md`), a detail link. A click inside a zone selects that zone, as
today. Everything else in the canvas box deselects: a shape of the drawing (bone, outline, any path, rect,
circle or group), the drawing's background, the box margin around the drawing.
- Drawing tools (point, zone, link) and a tracing: unchanged, a click places or adds a vertex.
- Gestures: the existing `swallowClick` of `useFraming` (a press that moved more than 4 CSS px: pan, marker drag,
pinch) already keeps the click from acting; in addition a press during which a second pointer came down (a
two-finger tap or a pinch that moved less than 4 px) never deselects, tracked in the canvas hooks so that
`useFraming`, shared with the catalogue reader, is unchanged. The wheel, Ctrl+wheel, the keyboard pan and zoom,
the zoom buttons and a touch scroll of the page (it ends in `pointercancel`) produce no click on the drawing.
- Effect: the selection and its contextual panel are reset (the overview comes back); document text, handle
version, history length and the save line unchanged; no save, no history entry. A field typed in the panel and
not blurred is committed by its own blur before the click, exactly as with « Toutes les annotations » today.
- Focus and announcement: the click leaves the focus on the drawing box it landed on (no jump to the list row,
which stays the rule of the « Désélectionner » button, A6); the same polite announcement as the button.
- Unit tests for D6b (`apps/app/tests/unit/editor-annotations-tab.test.tsx`, happy-dom; the test drawing has an inner
`<rect>`): with a selection, a click on the drawing's inner `<rect>`, on its root `<svg>` and on the box margin
each deselect, document text, handle version and history length unchanged; a click on another marker selects it,
on the selected one keeps it; nothing selected, a click on the `<rect>` changes nothing and a pending keyboard
creation keeps its typed name; a press moved more than 4 px on the `<rect>` (pan), a marker drag and a press with
a second pointer keep the selection; undo after a click deselect undoes the previous edit; the focus stays on the
box and the announcement is made; the read-only twin deselects on the `<rect>`; `CatalogueCanvas` without
`onEmpty` keeps today's behaviour.
- Unit tests: deselect leaves the document text and the history length unchanged; undo after a deselect undoes the
previous edit; the Escape matrix (dialog, menu, panel, tracing, nothing open) with one consumer each time; focus
after « Désélectionner »: the annotation's list row if it still exists, else the drawing, with no new selection
and a polite announcement (A6 as ratified); accessible names of both history buttons in FR and EN; read-only modes
show « Désélectionner » and no « Supprimer ».
- E2E: the journeys that name « Annuler » / « Rétablir » rewritten one for one; a new desktop spec for the button and
Escape; axe on Annotations with a selection. For D6b, in that spec (Chromium, `pnpm e2e:shell` against
`pnpm db:sandbox`, never the trial stack): real clicks on the inner `<circle>` and `<rect>` of the drawing written by
`e2e/support/editor.ts:78-80`, on the drawing's background and on the box margin, then on another marker; a drag
pan on the zoomed drawing and a marker drag keeping the selection; a touch context (`hasTouch`, as in
`editor-geometry.spec.ts:419`) where a tap deselects and a touch drag does not. `editor-equine.spec.ts`: one click
on a bone path of the historical horse drawing of the working copy deselects.
#### Behaviour and limits (as built)
- Owner decisions at the gates. G1: Escape moves the focus only when it was inside the properties panel (to the
annotation's list row if it still exists, else to the drawing); from anywhere else the focus stays where it was.
G2: the « Supprimer » button style is unchanged.
- D6b in plain terms: the rule follows the press, not the click target. A press that begins outside every
`[data-anno-id]` element and ends as a click deselects. A press that began on an annotation never deselects, even if
it ends elsewhere. A drag, a pan, a zoom, a pinch or a press that gained a second finger never deselects.
- Escape from outside the editor workspace (sidebar, application header) does nothing.
- Escape with the focus on the page (`body`) respects the layers above: an open modal dialog (delete, phone drawer)
keeps the selection, and while the references panel is open Escape closes the panel first and keeps the selection.
In « Plus », Escape typed in a text field or outside the workspace is left alone.
- A pen tap follows the touch rule: the decision is taken in the tap's own click, and only when that click's target is
not an annotation; only the mouse acts at the pointer-up. A hover move after a press is not part of it, and a new
press by the same pointer outside the box cancels a stale one.
- The read-only twin keeps its keyboard pan and zoom (arrows, +, -, 0) when it deselects.
- A value refused as « non appliqué » in the properties panel is lost when the panel closes, as with « Toutes les
annotations ».
- « Toutes les annotations » keeps its focus behaviour: when its row is filtered out, the focus falls to the page.
- The vertex handles of a selected polygon are display-only: a press on the part of a handle outside the zone's fill
deselects.
- A touch tap whose click Chromium drops does not deselect; a second tap does. This is rare: a field blur commit
rebuilds the drawing between the synthesized mousedown and the click.
- Chromium touch adjustment can turn a tap about 6 px outside a marker into a selection of that marker (no deselect
announcement then).
- Proof of the touch guards: Chromium only, through CDP touch with a capture-phase event recorder (identical over three
runs, `.superpowers/sdd/2026-11-12-r1-history-deselect-1141/task-5a-report.md`). A tap on empty: 1 pointerdown, 1
pointerup, 0 pointercancel, 1 click, deselect. A one-finger drag on a movable point: 0 click. A two-finger tap and a
pinch: 1 non-primary pointerdown, 2 pointerup, 0 click, selection kept. A one-finger drag on the read-only twin: 1
pointercancel (the browser pan took the touch), 0 click. iOS Safari is unverified: a manual check is listed in the
TODO, it is not a proof.
- The « Sélection : … » tag is not shown (screen-reader-only text, not live): the highlight and the properties panel
carry the selection, and the drawing keeps its place when the selection changes (owner, 2026-11-12).
### R2 — The Références tab (D1 to D4)
Status (2026-11-15, closure): VALIDATED by the owner, functionally and graphically, after the `db:trial` checks of
§4.13 and §4.14; R2 is closed (« Owner validation and closure 2026-11-15 » at the end of this section). No merge, no
deployment.
Status (2026-11-15, later): four pre-closure adjustments built (« Pre-closure adjustments 2026-11-15 » below,
`V1-2-6_OWNER_TRIAL.md` §4.14), awaiting the owner's checks, then R2 closure.
Status (2026-11-15): built, awaiting the owner's `db:trial` check of two last adjustments (« Owner follow-up
2026-11-15 » below, `V1-2-6_OWNER_TRIAL.md` §4.13), then R2 closure. Several lines below describe the side sheet, the
citations in the finding's edit form or the phone tabs Annotations + Vérification: each is marked as superseded.
Status (2026-11-14): R2 validated by the owner's trial, closed pending the final manual checks (« Owner trial and
closure 2026-11-14 » below, `V1-2-6_OWNER_TRIAL.md` §4.12).
Status (2026-11-13): corrected, still awaiting the owner's validation (corrections below).
Status (2026-11-12): built on `feature/v1-2-anatomical-sheets`, awaiting the owner's validation (nothing merged,
nothing deployed). Commits: `65a8fd3` (plan, rev 4 after the owner's GO), `fc3fec7` R2.0 (the panel's place renamed
`citations`), `4499bc7` R2.1 (fifth tab, panel from a finding), `c21995b` R2.2 (collapsible tiles), `3baac4d` R2.3
(search, sort, count, add), `8db6f42` R2.4 (contextual panel, board 25), `4b7d254` R2.5 (non-regression sweep), `2182307` (documentation),
`1bf2701` (final review fixes: one history entry per reference edit, focus kept).
Labels: `V1-2-6_R2_LABELS.md`. Decision record: BDR-114. What the build settled against the text below:
- The panel opens from a finding only (owner, 2026-11-12). The family's former trigger is removed; a family keeps
its inline list and « Retirer ». Superseded 2026-11-15: a family opens the references dialog like a finding, and
neither holds an inline list.
- « Retirer de ce constat » removes that one citation (the reference stays in the document) and marks the finding
« À confirmer » (core rule, `edit.rs`); one « Défaire » restores the citation and the status together. Superseded
2026-11-14: the sheet no longer removes a citation (read-only, « Modifier » and « Aller à » only, key
`editor.references.remove_citation` deleted); a citation is removed from the finding's edit form (see « Owner trial
and closure 2026-11-14 »). Superseded 2026-11-15: the finding's edit form holds no citation; the references dialog's
own « Modifier » cites and removes (see « Owner follow-up 2026-11-15 »).
- Zero state: when the last citation is gone the panel stays open with « Ce constat ne cite plus aucune
référence. » and closes on a tab change. Superseded 2026-11-15: the references dialog also opens at 0 reference, and
the line reads « Ce constat ne cite aucune référence. ».
- Selecting a tab closes the panel. Every close (✕, Escape outside a field, scrim, tab, « Aller à », « Tout
afficher », element gone) records one pending move applied after the exit: the focus goes back to the trigger,
else to the active tab button.
- A read-only flip (window under 1 024 px) keeps the panel open, in its read view: implemented, pending owner
ratification (the contract's criterion 15 and P22'' still say the panel closes). A refused draft the flip drops is
announced (« Saisie non appliquée : … »).
- Motion goes through `motion/react` (slide in 180 ms, out 140 ms, none under reduced motion, callbacks still
fire), not through the Tailwind `motion-reduce:` text below: deviation approved by the owner, 2026-11-12.
- The Places area keeps a permanent `dvh` minimum height so that opening the panel never changes the document's
size; the panel fits the visible viewport at every scroll position.
- Limits: at 667 × 375 (landscape phone) and scroll 0 about 13 px of the panel show, because the trigger sits in
Places, itself below the fold, so a scroll comes first (the e2e now covers it by touch, see Corrections 2026-11-13). A control below a field
whose blur-commit is refused can lose its first click, the error line shifting the layout: an editor-wide
pattern, not a regression, left out of R2. `admin.spec.ts` crashes during axe (« Target crashed ») under the
accumulated sandbox grants, not an R2 regression (`BROWSER_TESTS.md` §6). The add draft's cancel control was added by
C2 (Corrections 2026-11-13).
- Gates at `1bf2701`: typecheck, lint, `test:run` (app 1510), build, `e2e:ui` 251; `e2e:shell` 16 specs
(`admin.spec.ts` and `editor-equine.spec.ts` excluded): 117 passed, 108 skipped, 1 failed (`crossing:211` stuck on
the MFA step-up screen before the editor; 11 passed when rerun alone), full-row guard clean.
- Gates at `4b7d254`: typecheck, lint, `test:run` (app 1488), build, `e2e:ui` 251; `e2e:shell` sweep of 21 specs
without `editor-equine`: 219 passed, 116 skipped, 3 failed (all `admin.spec.ts`), full-row guard clean.
`editor-equine.spec.ts` was updated for typecheck and never run.
Corrections 2026-11-13 (owner brief), still awaiting the owner's validation (nothing merged, nothing deployed), on
top of `52aac41`:
- C1 `3303530`, `1e19479`: a refused draft is never dropped silently on a read-only flip. The tile in edit stays
shown through the flip, a drop is said once in the live region and shown in its tile, a forced close of the panel
names every refused field once, and the flip makes no history entry.
- C2 `d177992`, `c8109af`: « Annuler l'ajout » on the draft tile, and Escape from non-text targets and from a
pristine text field (never during IME composition). The document and undo/redo are untouched; the focus goes back
to « Ajouter une référence ». `c8109af` also hardened a flip unit test: happy-dom 20.10.6 holds MutationObserver
callbacks in a WeakRef, so a garbage collection under load stops the observation; the test now counts writes to
the live region.
- C3 `b1fd896`: a pinned tile that no longer matches the search shows « Ne correspond pas à la recherche » under
its header (`aria-describedby`). « Aucun résultat » shows only when no tile is shown (the add draft does not
count).
- C4 `969a5f3`, `817f74b`, `7766929`, `f7f0aa6`, `78fb2d0` (test only): proof, no app code. One test of
`editor-shell-layout` (phone project, 667 × 375), two parts. (A) Touch only: CDP `Input.dispatchTouchEvent` swipes
of 200 px from scroll 0 (`Input.synthesizeScrollGesture` with `touch` scrolls nothing in this Chromium) until the
trigger is wholly on screen and hit-testable (2 swipes in every run, scrollY 491 to 512); a tap opens the panel,
heading and ✕ in the viewport and hit-testable, scrollY unchanged. (B) Worst case: an instant programmatic scroll
puts the trigger's bottom on the viewport bottom (gap 0, scrollY 480); a tap opens the panel, heading and ✕ in
view and hit-testable, scrollY unchanged. B is programmatic because a touch swipe under about 60 px scrolls
nothing (slop) and the gain varies under load, so a pixel-exact touch landing is not reproducible (an earlier
calibrated version oscillated by 55 px in the full sweep).
- Final review `94f8a8b`, `346a8fd`: with two or more refused fields, a user leave (« Terminer », the panel's ✕,
Escape or scrim, another tile's « Modifier », « Ajouter une référence ») now leaves the focus on the first refused
field with its own error shown. Before, the last field tried owned the single error message (a regression of C1),
and a mouse click on « Terminer » with Citation blank and a valid URL showed no error at all (present before R2's
corrections). The forced close still names every refused field once.
- New keys (provisional EN): `editor.references.add_cancel` fr « Annuler l'ajout », en « Cancel adding »;
`editor.references.outside_search` fr « Ne correspond pas à la recherche », en « Does not match the search ».
Rows in `V1-2-6_R2_LABELS.md`.
Limits kept after the corrections: undo/redo replacing a pending draft in a mounted view; two drops in one flip
speak only the last;
the rotation to 667 × 375 with the panel open at scroll 0 keeps its cramped exception (not built, an option for the
owner). An undo or redo that removes a reference whose tile holds a refused draft now names the dropped
fields once in the live region, from the drop itself (V1); a full page reload is out of scope (the page is gone, the
local recovery copy rules apply).
`admin.spec.ts` diagnosis (C0, 2026-11-13, sandbox only, guard OK on every run, no grant pagination, no reset, no
deletion):
- `public.anatomical_catalog_grants` held 2 269 rows at the start (by day: 2026-11-08 234, 11-09 11, 11-12 1 896,
11-13 128), `auth.users` 2 871. Each run adds 25 grants and 25 accounts.
- Interleaved A/B, `admin.spec.ts` alone (24 listed, 8 workers): the pre-R2 worktree at `65a8fd3` (same lockfile)
against HEAD `52aac41`, four runs (abH1, abB1, abH2, abB2). Each gave 13 passed, 8 skipped, 3 failed, with the same
« Target crashed » inside `expectAccessible` (axe): A0 « the rights screen with grants » (`/app/admin/access`) in
one locale, « admin screens » desktop and 375 in one locale. Host memory stayed at 11.3 to 11.6 GiB available.
- Dose-response in a quiet window (no other Playwright run): a temporary copy of the spec routed
`list_catalog_grants` to its 753 newest rows. Capped (2 526 grants in base): 16 passed, 8 skipped, 0 failed. Full
(2 551): 11 passed, 8 skipped, 5 failed, 4 of them « Target crashed ». Capped again (2 570): 16 passed, 0 failed.
The temporary copy was deleted, never committed.
- Conclusion: the accumulation of e2e grants in the persistent sandbox causes the failures (the rights screen
renders every grant under axe). They are identical before R2 and on R2: not an R2 regression. Still failing on
the accumulated sandbox: A0 « the rights screen with grants » and « admin screens » (fr or en, desktop and 375,
the locale varies from run to run). The cure, a paginated grant list, belongs to RR-13; blocker BLK-015 stays open.
#### Finition 2026-11-13
Status: built on `feature/v1-2-anatomical-sheets` on top of `36c6eb3`, awaiting the owner's trial (nothing merged,
nothing deployed). Source: the owner brief « finition R2 » (items 1 to 9) and the plan
`.claude/tasks/plans/2026-11-13-r2-ux-finish-1624.md` (R2F-1 to R2F-9). Labels: `V1-2-6_R2_LABELS.md` §5 to §7.
Trial: `V1-2-6_OWNER_TRIAL.md` §4.11. Browser proofs: `BROWSER_TESTS.md`.
Commits:
- Task 1, dialog primitives and editor history: `553a0e5`, `69e225c`, `792394d`, `8e4f026`.
- Task 2, finding dialog and citations side sheet: `18085ce`, `4731a33`, `c893dec`, `b703333`, `20905ad`, `af300de`,
`e2d020b`.
- Task 3, reference dialog, read-only tiles, filter by finding, owner journey: `8031e00`, `7a64daf`, `8b42816`,
`3bb3603`, `8a60128`.
- Task 4, « Plus » overlay and sticky sidebar: `ca5804a`, `dfc8965`.
- Owner trial adjustments of 2026-11-14 (Tasks 6 to 8): see « Owner trial and closure 2026-11-14 » below.
Behaviour:
- Dialogs. A finding opens in a modal dialog (view mode, then « Modifier »); a reference opens in its own dialog (edit
or create). Forms are transactional: « Enregistrer » writes one history entry (nothing when pristine), « Annuler » or
any close with a change asks « Abandonner les modifications ? » first. A save is one « Défaire ». Reference tiles
are read-only (disclosure, « Modifier », « Supprimer » or the cited hint). Focus returns to the opener.
- Side sheet. The citations panel is a modal sheet at the right edge (34 rem at most) stacked over the finding dialog,
usable by keyboard and touch. Content: read tiles, « Modifier », « Aller à », « Tout afficher » (« Retirer de ce
constat » was removed from it on 2026-11-14). Superseded 2026-11-15: a modal references dialog stacked over the
finding dialog replaces the sheet (no sheet, no slide); its read tiles have no « Modifier ».
- History. Tab, Annotations section, finding dialog, finding creation, sheet, reference dialog, filter are browser
history entries (`apps/app/src/editor/history`). Back and Forward restore the view and the scroll; a reload keeps
tab and filter, never a form. Entries carry the sheet id and a random session tag (2026-11-14: no longer the account id) and hold
navigation only; they are reset on a sheet or account change. A Back that would drop a dirty form re-pushes the view and shows the confirmation.
- Filter. « Tout afficher dans l'onglet Références » shows only the references the finding cites, with the chip
« Références du constat : <nom> » and its « × », plus « Revenir au constat ». Search and sort stay usable. The
owner's eight steps (finding A, « 2 références », sheet, « Tout afficher », filtered tab, Back, back on A with the
sheet open) are proved at 1280 and 375 px (since 2026-11-15 the sheet of these steps is the references dialog).
- « Plus ». The menu is an overlay under the tab row (no layout shift, light fade), closed by a choice, an outside
press, Escape, a tab change and, where the browser offers `CloseWatcher`, the Android Back request. Superseded
2026-11-15: a vertical list of full-width rows (see « Owner follow-up 2026-11-15 »).
- Sidebar. The left menu stays in place while the page scrolls (own scroll when taller than the window).
Owner decisions (gate 2026-11-13): R2F-9 items 1 to 6 approved. Item 7: the clients and animals list searches are NOT
kept in browser history (BDR-078 stands). Precisions: one undoable operation per save; the side sheet usable by keyboard
and touch; history keeps navigation state, never typed data, personal searches or drafts; history reset on a sheet or
account change; the Android Back request closes « Plus » first; no new mockups; the limits below are accepted
provided no user change is lost silently.
Stated limits:
1. Deleting a finding replaces its dialog entry, so the next Back press does nothing visible once (R2F-3). Lifted
2026-11-14: the deletion pops the dialog entry, the first Back moves.
2. Lifted by the final review: « Aller à » a finding or a zone opens its dialog in the same move, one history entry;
Back returns to the place left, references dialog included (the sheet until 2026-11-15).
3. A move made in the same instant as a pending Back can be swallowed (not reachable at human speed).
4. « Aller à » a finding opens its dialog in view mode (accepted again by the owner, 2026-11-14).
5. The search of the References tab is not in the history: a plain tab change keeps it, « Tout afficher » clears it;
Back still restores the filtered view.
6. Android Back closing « Plus » relies on `CloseWatcher` (Chrome Android). Proved by a stub test only, not on a
device. Without `CloseWatcher`, Back navigates as before; Escape and an outside press still close.
7. A tap outside the open « Plus » menu only closes it (the page does not receive that tap).
8. After choosing a « Plus » entry by keyboard, focus goes to the page body (pre-existing). Lifted 2026-11-14: the
focus goes to the new tab's panel.
9. BLK-015: `admin.spec.ts` still fails on the accumulated sandbox (grant accumulation; pagination is RR-13).
Unchanged by the finish.
Also stated in the plan: leaving the editor by Back past its first entry cannot be refused (BDR-078, recovery copy).
A dirty dialog lost that way (a Back of several entries from the history menu) is no longer silent: one flag in this
tab's sessionStorage (layer and element key, never the text) makes the next mount of that sheet for that session tag say
once « Une saisie non enregistrée a été abandonnée en quittant l'éditeur »; the text itself is not kept. Since
2026-11-14 the browser is asked first where it can cancel the jump (see the closure subsection); the flag and the
notice remain the fallback.
Chromium's skippable-entry intervention and Android's close-watcher grouping of two restored dialogs are untested
(Playwright's `goBack` bypasses them) and are checked in the owner's phone trial.
Gone with the finish, so struck from « Limits kept after the corrections » above: the cramped view after a rotation to
667 × 375 with the panel open at scroll 0 (the panel is now a modal sheet at full height); the two refused drops spoken
as one (`draft_dropped` for references and the inline drafts are removed); a pending draft lost by undo or redo in a
mounted edit view (no inline draft remains, « Défaire » and « Refaire » are unreachable while a dialog is open). Still
true: a button under a field whose blur-commit is refused can lose its first click (other inline editor forms).
Status: validated by the owner's trial on 2026-11-14 (next subsection). Not run on a device: Android Back with a real
`CloseWatcher`.
#### Owner trial and closure 2026-11-14
Validation: the owner's trial of 2026-11-14 validated the R2 main journeys (`V1-2-6_OWNER_TRIAL.md` §4.11). The
trial asked for last adjustments, built on `feature/v1-2-anatomical-sheets` on top of `afc1051` (plan « Rev 4 »,
`.claude/tasks/plans/2026-11-13-r2-ux-finish-1624.md`, contract criteria 19 to 25). Nothing merged, nothing deployed.
Adjustments:
- References are cited only in the edit mode of a finding (`f0e49fd`, `270444e`). The view mode of the dialog only shows
the « n références » button that opens the sheet (a family shows its list read-only). The edit form keeps a draft list of the citations with « Retirer » on each and « Citer
une référence »; the one « Enregistrer » writes the citations with the other fields as one undoable operation. The
citations sheet is read-only for the finding (no « Retirer de ce constat »); while the form is open it lists the
draft's references. The key `editor.references.remove_citation` is deleted (`V1-2-6_R2_LABELS.md` §9).
Superseded 2026-11-15: the finding's edit form no longer holds citations (the references dialog's own edit mode
does), and the sheet is a dialog.
- Deletion and Back (`085c568`, `4018344`): the deletion closes its dialog by popping the entry, so the first Back
moves; the first « Défaire » after a deletion restores the finding at once (Constats opens, the row is selected and
focused). Choosing « Plus » entry puts the focus on the new tab's panel (the panels take `tabIndex=-1` and show a
ring only on keyboard focus, with a transparent outline for forced colours).
- Session tag (`07bebbc`, `579d8af`): history entries carry a random 128-bit tag kept in `sessionStorage`, never the
account id. The tag is bound to the account by a non-secret fingerprint (FNV-1a 64, synchronous, so it works on the
http LAN origin where `crypto.subtle` is missing), rotates on another account and at sign-out, and the old
left-draft flags are cleared. A record of the earlier shape, or an empty tag, reads as foreign (fail closed).
`AuthProvider.tsx` starts the sign-out subscription (3 lines, outside the first file scope, amended by the
controller).
- Leave confirmation (`07bebbc`, `579d8af`, `791fc03`): where the browser exposes the Navigation API and the event is
cancelable, a jump of several entries out of the editor with a dirty finding or reference dialog is cancelled and
the top dirty dialog asks « Abandonner les modifications ? ». « Abandonner » runs the jump again with
`traverseTo`, leaving no flag and no notice. With two dirty dialogs the question cascades from the top. Before
the finding asks, the citations dialog (a sheet until 2026-11-15) and a clean reference dialog over it close, so its
question is never hidden under them. Without the API or when the event is not cancelable, the earlier flag and notice stay.
- Notice after the loss elsewhere: unchanged text, now keyed by the session tag.
- Final review, rev 4 (2026-11-14): a Back asks only the layers it removes. A Back that closes only the citations
sheet, or a clean reference dialog over it, closes it and asks nothing, even over a dirty finding form (since
2026-11-15 a dirty citations draft is asked like the finding form). A Back or a
jump that removes the dirty finding dialog first closes the clean layers over it, so its question is visible and
focused. A sign-out, or a session ending, while a dialog holds unsaved text leaves one flag with no text, keyed by
the sheet id and by a fingerprint of the account salted with a random value stored in the flag: the same account's
next sign-in in this tab shows the notice once on that sheet, another account never matches. The session record's
fingerprint is salted with its own tag; a record of the older, unsalted shape reads as none and rotates.
Support of the confirmation (what `task-8-report.md` of the run proves): verified on Chromium desktop (the pinned
Playwright Chromium, a history jump through the DevTools protocol, one dirty dialog and two stacked); the fallback
logic is covered by unit tests (no API, non-cancelable event). Not verified: Firefox, Safari, a click in the real
toolbar history menu (editing needs 1 024 px, so a phone never reaches this case). The http-origin restore is proved by a unit test only.
Accepted limits (owner, 2026-11-14):
1. A move made in the same instant as a pending Back can be swallowed (not reachable at human speed).
2. « Aller à » a finding opens its dialog in view mode.
3. The search of the References tab is kept across tabs (not in the history; « Tout afficher » clears it).
4. The first tap outside the open « Plus » menu only closes it.
Limits found during these adjustments, to be confirmed by the owner at the final checks (`V1-2-6_OWNER_TRIAL.md`
§4.12):
5. In the leave cascade, « Continuer la saisie » on a lower dialog keeps that draft; a draft already abandoned above
stays dropped (the user's choice). The references dialog (a sheet until 2026-11-15) and a clean reference dialog
over the finding close before its question; one Back then reopens the references dialog.
6. Android Back closing « Plus » relies on `CloseWatcher` and is proved by a stub test only.
7. BLK-015 and `admin.spec.ts`: tracked apart (RR-13), unchanged.
Lifted by this round: the dead Back after a deletion (finish limit 1) and the focus lost to the page body after a
« Plus » choice (finish limit 8). Changed: the account id left the history entries for the session tag; a multi-entry
leave is confirmed where the browser can cancel it.
Status: R2 validated, closed pending the final manual checks (`V1-2-6_OWNER_TRIAL.md` §4.12). The owner still has to
check Android Back on a real phone.
#### Owner follow-up 2026-11-15
Status: built on `feature/v1-2-anatomical-sheets` on top of `b41c24b` (plan « Rev 5 »,
`.claude/tasks/plans/2026-11-13-r2-ux-finish-1624.md`, contract criteria 26 to 31), awaiting the owner's `db:trial`
check (`V1-2-6_OWNER_TRIAL.md` §4.13), then R2 closure. Nothing merged, nothing deployed.
The trial validated the finding dialogs and « Défaire » / « Refaire ». Two adjustments were asked.
Requests (owner brief 2026-11-15):
1. References apart from the finding. The finding dialog keeps « Références citées » and « X références » in view and
in edit mode. The button opens a references dialog stacked over it, which replaces the side sheet. View mode only
reads. Its own « Modifier » cites an existing reference, removes one, creates one (cited in the draft), opens a
reference's dialog, saves (one « Défaire ») or cancels. Families get it too.
2. One tab order: Général, Annotations, Constats, Références, Vérification. On phones Général and Annotations are
shown and the rest sits under « Plus », in that order. « Plus » is a vertical list: rows of 48 px at least, the
active entry tinted with a check mark, light motion, the error count on « Plus » while Vérification is under it.
Built, by task:
- Task 10, one tab order and the vertical « Plus » list: `ff21b03`, `da755f9` (hover tint, axe on an active row, line
limits). `tabs.ts` shows the first two tabs on phones; `initialTab` lost its tier parameter. The active row carries
`aria-current="true"` besides the tint and the check mark.
- Task 11, the references dialog in view mode: `abc781d`, `2c4c771`, then `d9dbcf1` with the review minors. The trigger
« X références » is a dialog trigger (`aria-haspopup="dialog"`). `CitationsSheet`, its motion, `FindingEditCitations`
and the shortcut store `citationDraft.ts` are gone. The finding's edit form no longer carries `reference_ids`: its
save never rewrites citations.
- Task 12, the dialog's edit mode: `3dc5820`, `2579d4c`. Draft held as adds and removes, applied to the element's
current list at save, so a reload or a change made elsewhere is never overwritten silently. « Enregistrer » writes
one `update` of `reference_ids` (none when pristine), then back to view mode with « Modifications enregistrées » and
the focus on « Modifier ». A new guarded layer `citations` sits between `finding` and `reference`: a dirty draft
refuses ✕, Escape, backdrop, Back and the leave cascade like the finding form.
- Task 13, this documentation.
Controller rulings (autonomous mandate, reported to the owner):
- Families (categories) get the same dialog as findings; both hold citations.
- View mode is read only: no « Modifier » on its tiles; « Modifier la référence » lives on the edit form's rows.
- « Nouvelle référence » creates a reference and cites it in the draft.
- On phones the visible tabs are the first two of the order; no arrow-key navigation in « Plus » (Tab moves, Tab out
closes).
- A reference created from the dialog, then the draft abandoned, stays in the library uncited. The creation is its own
undo step, tested and documented. Superseded 2026-11-15 (pre-closure): the creation stays in the draft until
« Enregistrer » and leaves nothing in the library after an abandon.
Labels: `V1-2-6_R2_LABELS.md` §10. Browser proofs: `BROWSER_TESTS.md`. Gates at the last Task 12 commit: typecheck and
lint 0, `test:run` 1683, `e2e:ui` 251, the touched `e2e:shell` specs green with the full-row guard clean. The full
`e2e:shell` sweep is not rerun for Rev 5 (BLK-015 `admin.spec.ts` unchanged).
Status: awaiting the owner's `db:trial` check (`V1-2-6_OWNER_TRIAL.md` §4.13), then R2 closure. Superseded in part the
same day: see « Pre-closure adjustments 2026-11-15 » below.
#### Pre-closure adjustments 2026-11-15
Status: built on `feature/v1-2-anatomical-sheets` on top of `ae02f48` (plan « Rev 6 », contract criteria 32 to 36,
`.claude/tasks/plans/2026-11-13-r2-ux-finish-1624.md`), awaiting the owner's checks (`V1-2-6_OWNER_TRIAL.md` §4.14),
then R2 closure. Nothing merged, nothing deployed.
The owner's trial of the Rev 5 build (§4.13) asked for four more adjustments before closure.
Requests (owner brief 2026-11-15):
1. A « Citer » button. Choosing a reference in the select no longer cites it: the choice only fills the select, and
« Citer » (`type="button"`) adds it to the draft, then resets the select. Arrow keys on the select and closing with
only a choice ask nothing.
2. A reference created from the edit mode belongs to the draft. It is written with the citations by « Enregistrer »,
in one undo step, and an abandon leaves the library as it was.
3. Family wording. A family dialog and the places that name its element say « famille », not « constat ».
4. A close button easy to touch. The ✕ of every dialog is a 44 px box.
Built, by task:
- Task 14, « Citer », family labels and the ✕: `240c2ba` (the ✕), `8c3fcca` (« Citer » and the family labels, one
commit: the fr, en and e2e hunks are shared), `1d119c3` (review round: a stale choice resets at render, the pending
line sits under the add row).
- Task 15, the creation in the draft: `70eb2fc`, `ee79127` (review round: the notice waits until the reference dialog
is gone, a locale-aware list of pending references, a note when a pending reference is edited).
- Task 16, this documentation.
Rulings (controller, after a three-lens challenge of the plan):
- « Enregistrer » with a chosen reference not yet cited refuses: the form stays in edit mode, a line at the select says
so (`editor.citations.pending_choice`) and the focus goes to « Citer ». The select value is derived at render from
the available options, so a choice whose reference vanished goes back to the placeholder.
- Family rule: one helper `isFamilyKey(key)` (a finding key with a one-segment path), beside `isCitable`, drives every
« famille » variant: the dialog titles, the count and empty lines, the citer row (« cette famille » alone, no second
badge), the filter chip, its remove and back labels, the vanished-filter announcement and the « n'existe plus »
line. Plural keys keep the unchanged plural helper.
- The pending reference travels from the reference dialog to the citations form through the layer's held state, never
the navigation state or the history. « Modifier la référence » on a pending row reopens it prefilled (title
« Nouvelle référence », « Valider »); « Retirer » on it drops it. The insert is dry-run on the current document text
before it is accepted, so a refusal shows in the reference form. The pending row carries a « Nouvelle » badge.
- The save writes one `insert` per pending reference then the `update` of the citations as one batch. A clashing id is
minted again once. A refused batch keeps the draft and names the pending references.
- Notices about a pending reference (created, edited, removed) go to the citations dialog's own status line, never to
the page's live region, so an abandon leaves nothing announced.
- « Modifier la référence » on an existing reference keeps its own save: its dialog writes at once, as a library
change with its own « Défaire ». Only a reference created from the edit mode is deferred.
Superseded by these adjustments: the Rev 5 ruling that a reference created from the dialog stays in the library
uncited after an abandon; the select that cited on change; the dialog ✕ at its icon size (21 px wide in Chromium).
Limits:
1. A pending reference lives only in the open references form until « Enregistrer ». The leave guards (✕, Escape,
backdrop, Back, the leave cascade) and `beforeunload` protect it, but no recovery copy is kept: a reload, a tab
closed or a killed tab loses it, as for the other drafts of this dialog.
2. « Modifier la référence » on an existing reference is a library change at once, not part of the citations draft. It
cannot be abandoned with « Annuler » of the citations form; its own « Défaire » removes it.
3. The English wording is provisional, as for every key.
4. `useEditCommand.check` dry-runs on the document text without the store's blocked state; the form is read-only
then, so its submit is off.
5. Not seen red: the new shell cases (written with the code), the alignment assertion of « Citer » and the unit case of
the unreadable element at save (it pins existing behaviour).
Labels: `V1-2-6_R2_LABELS.md` §11. Browser proofs: `BROWSER_TESTS.md`. Gates at `52d5613`: typecheck and lint 0,
`test:run` 1722 (app), `e2e:ui` 253, full-row guard clean on every `db:sandbox` run. The full `e2e:shell` sweep is not
green. At `ee79127` (final6): 269 passed, 162 skipped, 9 failed: `admin.spec.ts` 6 (BLK-015),
`editor-shell-layout.spec.ts:182` and `:245` (« Target crashed »), `animals.spec.ts:664` (page load timeout at
375 px); the last three passed when their two specs ran alone. At `52d5613` (final7): 273 passed, 163 skipped,
6 failed, all `admin.spec.ts` (BLK-015).
Original plan text:
- Tab model: `EDITOR_TABS` gains `references` between `findings` and `verification` (`apps/app/src/editor/tabs.ts:6-11`),
`SHOWN.small` keeps it under « Plus » (:18). The id collides with the panel's `data-place` name `references`: the
panel place is renamed first (internal name, no visible effect). Touched with it: `navigation.ts` `targetOfKey`
(:65-81, reference keys go to the tab), `focusRequest.ts:19-22`, `useEditorNavigation.ts`, `EditorTabBody.tsx:16-23`,
`EditorWorkspace.tsx:85-131`, `verification.ts`, i18n keys.
- Tab content: tiles built from `ReferenceRow` / `ReferenceForm` (`components/editor/references/`), a disclosure
button per tile (`aria-expanded`, `aria-controls`), the toolbar with a search field and its explicit clear button
(focus back to the field), a native `<select>` with its external « Trier par » label. Tile count reads « Citée N
fois » (citations from families and findings, as `citations.ts` counts them); the open tile names a family as a family.
A cited reference shows no « Supprimer » (the core refuses it): a hint explains how to delete it once each citation
is removed (« Aller à », then « Retirer » in the finding's or family's detail); an uncited reference keeps « Supprimer » (owner verdict,
board 23).
« Ajouter une référence » appends in document order, resets search and sort, opens the new tile in edit with focus
on its citation field. Sorting is a pure function over the reference ids that never
touches the document: document order; citation A to Z with `Intl.Collator` of the current language; citation count
descending, ties in document order; type in the order of the type list, ties in document order. Search keeps
`filter.ts` (citation, DOI, URL, `folded`).
- Panel: opened only from a finding (`EditorWorkspace.tsx:94` width becomes `min(34rem, 100%)`), slide from the right
in 150 to 200 ms with a shorter exit, no slide under `prefers-reduced-motion` (Tailwind `motion-reduce:`); the panel
stays mounted during its exit and the page stays `inert` until the exit ends (LRN-085); no permanent `will-change`
(LRN-045). The header button and its focus fallback are removed (`HeaderActions.tsx`, `REFERENCES_BUTTON_ID` in
`tabs.ts:57`).
- Unit tests (about 15 files today assert the four-tab structure or the panel): five tabs and the « Plus » tier at
500 px; each sort, stable, document untouched (text compared before and after); reorder commands present only in
document order without search; clear button; tile disclosure and focus; « Aller à » from Vérification opens the
tab, the tile open and focused; « Aller à » from a tile reaches the finding; a cited tile has no « Supprimer » and
shows the hint, an uncited tile keeps it; the panel from a finding: filtered, « Retirer de ce constat » removes only
that citation (the reference stays in the document) and one undo restores it, « Tout afficher » opens the tab with search and sort reset and focus on its heading, close returns focus to the
finding's button; unsaved edits in a tile form survive a tab change (kept-mounted places); no edit control in any
read-only mode or without the editor right; the reduced-motion class.
- E2E: `TABS` helper (`e2e/support/editor.ts:327`) to five; `editor-shell-journeys.spec.ts:116-125` count 4 to 5;
layout specs with the panel open at 1 024, 1 180, 1 279, 1 280 px (no horizontal scroll); 375 px « Plus »;
axe on the tab (tiles closed and open) and on the panel; keyboard walk of the new order; about 28 panel open/close
calls rewritten.
#### Owner validation and closure 2026-11-15
Owner brief « Clôture R2 » (2026-11-15): the manual checks on `db:trial` are done and pass. R2 is validated
functionally and graphically; no further correction is asked.
- The two choices offered with the pre-closure adjustments are ratified as built:
1. « Modifier la référence » on a saved reference stays outside the citations draft. A saved reference is
independent of the findings that cite it: its change applies to every citation and has its own « Défaire ».
2. A new reference not yet saved is lost if the browser is closed abruptly, like any other unsaved form.
- Test results at the closing code (`52d5613`), unchanged by the later docs commits: typecheck, lint and build exit 0;
`test:run` app 1722, supabase 180, core-wasm 66; `e2e:ui` 253 passed. Full `e2e:shell` sweep on the sandbox (without
`editor-equine`): 273 passed, 163 skipped, 6 failed. The six failures are `admin.spec.ts` (BLK-015, grant
accumulation, pagination in RR-13), which stays open as a separate task. The sweep is not green. The earlier sweep at
`ee79127` had three more failures under load (`editor-shell-layout.spec.ts:182` and `:245` « Target crashed »,
`animals.spec.ts:664` navigation timeout at 375 px); they passed when their specs ran alone and did not recur at `52d5613`, but they are not counted as fixed.
The full-row guard reported no change to pre-existing rows on every run.
- U2 orientation kept as recorded on 2026-11-14 (`V1-2-6_UX_PLAN.md` §U2 « Orientation 2026-11-14 »,
`V1-2-6_UX_DECISIONS.md` row « U2 orientation », `docs/design/README.md` note of 2026-11-14). U2 is not started.
Status: R2 closed. Next lots: R3, then R4 (below); U2 waits for its own planning and owner gate.
### R3 — Sheet list and Administration home (D7)
- `AdminSheetsRoute.tsx`: search on the sheet name in both languages (`folded`, accents ignored) with a clear
button, a « Trier par » select (last modification descending by default, name A to Z by `Intl.Collator`), the
result count, the date on each row; the four filters unchanged (`apps/app/src/admin/filters.ts`).
- `AdminHomeRoute.tsx` « Fiches officielles » card (`CARD_SHEETS = 5`, :25): the five most recently modified sheets
with species, status and date, and « Voir les N fiches ».
- Date source (arbitration A7): the list payload carries `anatomical_sheets.updated_at`
(`supabase/migrations/20261102100000_sheet_admin_reads.sql:54`), which a draft save does not move:
`save_sheet_draft` writes `anatomical_sheet_versions` only (`20261028110000_sheet_review_workflow.sql:569-630`). On
the sandbox, 28 of 219 official sheets have a version newer than the sheet row. A true « dernière modification »
needs one read change in SQL (a `last_modified_at` field, greatest of the sheet's and its versions' `updated_at`,
in `private.official_sheet_admin_summary`; no table, policy or privilege change), hence an SQL checkpoint, pgTAP,
regenerated types, and its application to a stack, which needs the owner's go for that operation. Ratified (A7):
the dedicated checkpoint is `V1-2-6_R3_SQL_CHECKPOINT.md`; R3 ships in two steps, search and name sort with no date
first, the date, the date sort and the « most recent » card only once the SQL change is authorised and applied.
Superseded 2026-11-16 by the owner's decision D2 (below): one delivery in the final state, the « before » state
never built.
- Search and sort run in the browser on the full list: `shortcut:` until RR-13 pages the list on the server, where
search and sort move with it.
- Tests: unit for search (accents, both languages, empty, clear), sort (date ties by name, collation), date format
FR/EN, the home card (at most five, sorted); e2e admin list search and sort at desktop and 375 px with axe; the four
filters unchanged.
#### As built 2026-11-16
Owner brief « R3 » (2026-11-16), plan `.claude/tasks/plans/2026-11-16-r3-sheet-list-0017.md` rev 2b validated the
same day: D1 to D4, D6 and D7 yes, D5 no, the « recorded unless the owner objects » choices accepted, precisions P1 to
P4 binding. Then the owner's go for `pnpm db:trial migrate`. Contract:
`.claude/tasks/contracts/2026-11-16-r3-sheet-list-1720.md`. Branch `feature/v1-2-anatomical-sheets`, no merge, no
deploy, DEV unchanged.
- Delivery (D2): one lot in the final state. The two-step split of A7 is superseded: no « before » state (search and
name sort without a date) was built or shown.
- SQL: M9 `supabase/migrations/20261116100000_sheet_admin_last_modified.sql` adds `last_modified_at` to
`private.official_sheet_admin_summary`, the newest of the sheet row's `updated_at` and of its versions'
`updated_at` (`V1-2-6_R3_SQL_CHECKPOINT.md`). Reviewed (task review, security review) before its first
application, frozen since (blob `a33b1770`). pgTAP `supabase/tests/022_sheet_admin_reads.sql`, plan 74: a saved
draft, a saved working copy, a sheet without a version, a sheet row newer than its versions (P1).
- Stacks (D3): the sandbox through the new `pnpm db:sandbox migrate` (additions only to `scripts/db-sandbox.sh`, no
reset), then the trial stack through `pnpm db:trial migrate` on the owner's go (cold backup
`20261116T144154Z-pre-migrate` first), then the application code. Figures: `V1-2-6_R3_SQL_CHECKPOINT.md` §6.
- Client (`packages/supabase/src/sheets.ts`): the summary requires `last_modified_at` as an ISO 8601 timestamp with
offset that a `Date` can read; an answer without it, with a value that is no timestamp, or with one the shape
accepts but a `Date` cannot read (`+99:99`, final fix wave), is the closed `unavailable` failure, never a list with
a missing or invented date (P3).
- List page `/app/admin/sheets` (`AdminSheetsRoute.tsx`, boards 28 and 29 A): search on the name in French or English
(`sheetNameMatcher`, the catalogue's rule folded once per search, beside its `matchesSheetName`: accents and case
ignored, surrounding spaces trimmed), instant, 100
characters at most, no magnifier; « Effacer la recherche » shown only while the field holds text; Escape empties a
non-empty field and does nothing on an empty one; every way of clearing returns the focus to the field. « Trier
par »: « Dernière modification » (default, newest first) or « Nom » (the collation of the interface language: French
names in the French interface, English names in the English one). Equal dates, then equal names, fall back to the
name, then the id, so the order never depends on the server's. The four filters unchanged, combined with the
search. A count line (`aria-live="polite"`): « {n} fiches », or « {n} fiches sur {total} » when something is
hidden. Each row is one link: name, the existing badges, the detail line (species, sheet status, published
version or none, draft version, latest round), and « Modifiée le 9 nov. 2026, 12:06 » in Paris time, inside a
`<time>` whose value is the same instant (`toISOString()`). No new badge: published and draft are told apart by
the detail line.
- Empty states (board 28 note, literally): no official sheet at all, « Aucune fiche officielle pour le moment. »;
nothing shown while a filter is active, the existing « Aucune fiche ne correspond à ces filtres. »; nothing shown
for a search without a filter, a plain block « Aucune fiche ne contient « {q} » dans son nom. » with the button
« Effacer la recherche ». The searched text is the trimmed text, inserted as typed (a `$&` stays `$&`), wrapping
anywhere.
- Administration home card (`AdminHomeRoute.tsx`, board 29 B): under « Brouillons, revues, versions. » the
sub-heading « Dernières modifications », then at most five sheets in the list's « Dernière modification » order,
each with its name link, badges, detail line and date. The historical sheet keeps its « Créer une copie de
travail » action on the card only when it is among the five, for an editor; otherwise the action is on its own
page (F9, A9). Footer link: « Voir les {n} fiches » from two sheets, « Toutes les fiches officielles » for zero or
one (D6). A failed species read leaves the species id in the detail line and never hides the card; a rights
refusal of the sheet list still hides it.
- Labels exactly as `docs/design/v1-2-6-r0/REVUE.md:82-84` in French and English; date format A8.
- Deviations from boards 28 and 29, stated in `docs/design/README.md`: no magnifier in the search field (F14, a
choice the owner validated with the plan); no info icon in the no-result block although board 28 draws one (a
ruling during the build, to be confirmed by the owner: `Icon` has no such glyph and the plan allows no new icon).
- Search, sort and the date format stay TypeScript, display only: rows in `RUST_ARCHITECTURE_REVIEW.md` Part II
(class C). The `shortcut:` of `apps/app/src/admin/sheetList.ts` (search and sort in the browser on the whole list,
until RR-13 pages it, D4) and the one of the search field in `AdminSheetsRoute.tsx` (same markup as the Références
tab's search, to fold into one component) are in `.claude/tasks/TODO.md`.
- Not in R3: the search, sort and filters kept when coming back from a sheet (D5, a later navigation improvement,
BDR-079 follow-up); R4 (board 30).
Results recorded task by task up to the last code commit `b4a5254` (exact; nothing below that failed is green):
| Check | Result |
|---|---|
| `pnpm typecheck`, `pnpm lint` | passed at each task's gate |
| `pnpm build` | passed at `423125c` and `de2424f` (the two screen commits) |
| `pnpm test:run` | core-wasm 66, supabase 181, app 1743 (at `de2424f` and `b4a5254`) |
| `pnpm db:sandbox test sheets-admin` (the client against the migrated sandbox's answers) | 5 of 5 |
| `pnpm db:sandbox test-db` (pgTAP on the real image, accumulated sandbox, after M9) | Files=21, Tests=1958, Result FAIL. `022_sheet_admin_reads` ok, `023` ok. `014_animals_relations`: 2 of 258 failed (tests 6 « exactly the nine species of the contract and other, equine unchanged » and 7 « every species row is active »). `021_sheet_verification_jobs`: 23 failed, « Bad plan. You planned 90 tests but ran 42 », an error `draft.frozen` at :205. Read as dependent on the accumulated sandbox state (the suites need a fresh stack); M9 touches neither file's functions. NOT proven: no run of these files without M9, and no fresh stack (`up fresh` needs an owner go) |
| `pnpm db:sandbox types --check` | no drift |
| `e2e/shell/admin-sheets-list.spec.ts` alone on the sandbox | 4 passed, 0 failed |
| Shell sweep, 27 spec files without `editor-equine.spec.ts` | 446 tests: 277 passed, 6 failed, 163 skipped, 0 flaky. All six failures in `admin.spec.ts`: « admin screens » (:714) ×4, fr and en at desktop and 375 px, `frame.evaluate: Target crashed` during axe at `admin.spec.ts:747`, inside that test's loop over the admin pages (the page is not recorded in this run; BLK-015's own description in `BROWSER_TESTS.md` names the rights screen as the proven cause in earlier runs); A0 (:220) ×2, on the rights screen, fr and en at desktop, the grant lookup stuck on « Recherche… » until the « E-mail confirmé » check timed out (:144), about 7 600 grant rows rendered. Counted as BLK-015, the A0 pair with another symptom than « Target crashed ». NOT proven: no A/B run against the commit before R3. Green as required: A1 and « access per right » (the rewritten locators, both languages), `editor-journey`, the new spec |
| Data guard (every pre-existing row of `anatomical_sheets` and `anatomical_sheet_versions`, per id) | 0 changed, 0 missing on every run; the 12 sheets the R3 spec created in its two runs read back `deleted` (D7) |
Closing gates on the range `98178de..73ed375` (2026-11-16):
- Design review (fresh, the two screens at 1 280 and 375 px on the sandbox, one Playwright run, 2 passed, data guard
0 changed, 0 missing): board 28 CONFORME (the two recorded deviations only); board 29 A ÉCARTS (2 small); board 29 B
ÉCARTS. Findings 1 to 4 are fixed in the final fix wave: 1, the home at 1 280 px stacked « Espèces » and « Accès et
habilitations » in the right column at their own height (`8bb2486`); 2, the card's sheet name links 44 px tall at
375 px (`8bb2486`); 3, « Effacer la recherche » of the no-result block 44 px tall at 375 px (`1758276`); 4, the
card item on the list row's rhythm (`8bb2486`). Left to the owner: finding 5 (below), the breadcrumb and the info
icon. Finding 6 (typography: an opening guillemet can end a line, « version / 1 » orphan in the detail line) is not
in the wave. Not checked by the design review: the badges and the working copy action on the card, the owner's
layout (deduced from the code), the English wording, the 0 or 1 sheet link, distinct dates, hover, screen reader
output.
- Final code review (fresh, opus): 0 Critical, 0 Important, 9 Minor, « Ready for the owner's trial: Yes ». Fixed in
the wave: 1 (each keystroke recomputed the whole list: the term folded once, the dates read once before the sort,
the rows memoised; no `useDeferredValue`), 2, 3 and 6 (`useListView` split, className constants, the
`SheetSummary` alias and the `ResultsProps` order, `1758276`), 4, 5 and 7 (the spec's `timezoneId`, the ordering
loop's count, the escaped pattern, Escape pressed in Chromium, `ea5512a`). Already done by the Task 8 documents: 9
(the `shortcut:` harvest). Left: 8 (the card checked for the true date: a flake risk on a shared sandbox).
- Security gate (fresh): PASS. semgrep `p/security-audit`, `p/secrets`, `p/owasp-top-ten`: 0 findings. Four
non-blocking findings: [MEDIUM] SQL built by template interpolation in the new spec, fixed (`literal()` and a uuid
shape check, `ea5512a`); [MEDIUM] guardrails API-PASSWORD-LITERAL at `scripts/db-sandbox.sh:485`, a pre-existing line
and a false positive, left; [LOW] the date checked by its shape only (`+99:99` passed and a `Date` read NaN), fixed
(`374e300`); [LOW] a RegExp built from unescaped dictionary text in the spec, fixed (`ea5512a`). Outside the diff:
the guardrails scanner reports 2 FAIL at `scripts/db-trial-check.sh:1340`, a file R3 never touched, not analysed
(TODO).
- Final fix wave (2026-11-16, `374e300`, `1758276`, `8bb2486`, `ea5512a`): unit gates core-wasm 66, supabase 182,
app 1743; `pnpm build` passed. One Playwright run of `admin-sheets-list.spec.ts`, `admin.spec.ts` and
`editor-journey.spec.ts`: 15 passed, 6 failed, 9 skipped. The new spec 4 passed; A1 and « access per right » green
in both languages; `editor-journey` green. The six failures, all `admin.spec.ts`, all on the accumulated sandbox:
« admin screens » (:714) fr desktop, en desktop and fr 375 px, `frame.evaluate: Target crashed` during axe at
`:747`; « admin screens » en 375 px, `page.evaluate: Test timeout of 180000ms exceeded` during axe at `:747`, the
screenshot on the rights screen; A0 (:220) fr, the « E-mail confirmé » check at `:144` after the lookup; A0 en, the
« Right granted. » check at `:158`, the button stuck on « Granting… ». Counted as BLK-015, not proven by an A/B run
against the commit before R3: the wave changed no file of the rights screen. Data guard: 0 changed, 0 missing, 20
new lines (10 sheets with their 10 versions); the 6 sheets the R3 spec created read back `deleted`.
- Verifier on the contract (fresh, 2026-11-16, at `bcdacbb`): ECARTS(1). Criteria 1 to 3 and 5 to 18 met.
Criterion 4 not met on its CHECK line only: the line greps a one-line schema literal that the fix wave (`374e300`)
split over four lines and made stricter. A layout-insensitive equivalent prints `DATE_SOURCE_OK`; it is recorded
in the contract. The owner closes criterion 4.
- `pnpm check:all` (2026-11-16, at `bcdacbb`): exit 0, « all checks passed » in 7 min 52 s. core-wasm 66,
supabase 182, app 1743 tests in 137 files, e2e:ui 253 passed, the application image, the worker tests and the
worker runtime image. e2e:shell is not part of `check:all`: its results are the sweep and the fix-wave run above.
- Still pending: the owner's trial on the trial stack (`V1-2-6_OWNER_TRIAL.md` §4.15). Not checked by the build: a
browser sign-in on the trial stack (the owner's accounts).
Owner trial 2026-11-16 (db:trial): general behaviour validated; one last UX request for the phone, built the same
day (`9efc7b7`, spec `6d2ba0d`), no new board:
- Below `sm` (640 px, the breakpoint the page already uses) the four filters sit in a collapsible panel, closed by
default, opened by a « Filtres » button (`aria-expanded`, `aria-controls`; 44 px; « Filtres (n) » with n the
non-empty filters). Search and sort stay in view. Filters apply at once, closing keeps them. « Réinitialiser les
filtres » sits inside the panel, phone only, shown while a filter is active; the focus then goes to the « Filtres »
button (the reset button disappears). From `sm` the grid stays as before: no button, no reset.
- Mechanism: CSS only, no viewport detection. The panel is a grid whose row goes from `0fr` to `1fr` (200 ms, none
under `prefers-reduced-motion`), with `visibility: hidden` while closed so nothing in it takes focus or a tap.
The state is `data-open` plus classes, unit-tested in happy-dom; the real behaviour is in the sandbox spec at 375 px and
1 280 px, fr and en.
- Double announcement (earlier open decision 5): confirmed in the code. With a filter and no result, the count line
(`aria-live="polite"`) and `role="status"` of the filters sentence both spoke. The sentence is no longer a live
region (the count line announces once, as `SearchNoResult` already did); the visible text is unchanged. The
sentence for « no official sheet at all » stays a status: no count line shows then.
- Verifier gap (criterion 4 of the contract): the one-line literal oracle was defective after `374e300` split the
line; the contract now holds the layout-insensitive check, run: DATE_SOURCE_OK. Criterion 4 met.
- Review: one review and two scoped fix rounds (`7a129dc`, `ca70178`, `a50796b`): toggle label left and chevron
right, closed panel of height 0 (asserted at 375 px), one `activeFilterCount`, owner message verbatim in the
contract. `pnpm check:all` at `a50796b`: exit 0, units core-wasm 66, supabase 182, app 1750, e2e:ui 253 passed;
e2e:shell not run by check:all. Spec admin-sheets-list alone: 8 passed. BLK-015 and the `admin.spec.ts`
failures stay in their own tracking, untouched.
Owner validation 2026-11-16: R3 validated officially after the owner's db:trial trial, the phone filters panel
included; « Réinitialiser les filtres » stays inside the panel only. R3 is closed. The points below were raised
before that validation; no change was requested on them, so they stay as built.
Points raised before the validation (R3, as built):
1. The info icon of the no-result block: board 28 draws one, the page shows none (no glyph in `Icon`).
2. At 375 px « Nouvelle fiche » sits beside the title, where board 29 A draws it under (header older than R3).
3. The top-bar breadcrumb « Administration › Fiches officielles » the boards draw (shell level, older than R3).
4. With a filter and a search both active and no result, the page shows the filters sentence without « Effacer la
recherche » (board 28's note, followed literally).
5. In that same case, the double announcement: the count line and the `role="status"` sentence. Fixed (see the owner
trial entry above).
### R4 — Integrity warning in catalogue cards (D9 C, independent fix)
- The card already runs the integrity check: `useCardDetails` calls `loadVersion` and drops a failed load in
silence (`apps/app/src/catalogue/useCardDetails.ts`, `if (!load.ok) return`). R4 keeps the failure kind in the
card's state and shows a warning inside the card; the sheet page's `CatalogueFailureNotice` is unchanged. No SQL,
no Rust: the check stays the core's `check_published`, never bypassed.
- Tests: unit for a card whose load fails with `fingerprint_mismatch` (warning shown, link kept) and with a transport
failure (its own line and « Réessayer », arbitration A10); e2e on a catalogue entry whose stored fingerprint differs, built by
the test itself (never by editing an existing sheet); axe on the list.
- Its own commit and verification, kept out of the U1 shell lots.
#### As built 2026-11-17
Owner brief 2026-11-16 (R3 closed, R4 start), plan `.claude/tasks/plans/2026-11-16-r4-card-integrity-2337.md` rev 3
validated 2026-11-17 (« Je valide le plan R4 avec les décisions suivantes »): D1 alternative, D2 to D4 as recommended,
R-1 to R-8 accepted, precisions P1 (the trial tells the sandbox's anomalies apart from R4) and P2 (the spec cleans
only what it created) binding. Contract `.claude/tasks/contracts/2026-11-17-r4-card-integrity-0004.md`. Branch
`feature/v1-2-anatomical-sheets`, no merge, no deploy, DEV unchanged, no write on the trial stack. Board 30
`docs/design/v1-2-6-r0/30-catalogue-carte-integrite.png`, wording A10.
- The check is untouched: the card still runs `loadVersion` (version, drawings, `CataloguePreview`, the core's
`check_published`). Nothing changed under `crates/`, `packages/core-wasm/`, `catalogue/catalogue.ts`,
`catalogue/preview.ts`, nor on the sheet page (`SheetRoute`, `CatalogueFailureNotice`, `useSheetVersion`).
- `useCardDetails` returns `{ state, retry }`: `loading`, `ready` (date, counts, thumbnail, `viewRefused`) or
`failed` (the closed kind, `retrying`, the attempt number). `cardFailure(failure, stopped)` in
`catalogue/failure.ts` sorts the kinds; `components/catalogue/CardFailure.tsx` draws the line. A failed card keeps
the name, « Publiée », « Version n » (no date: the date comes with the details that failed) and « Consulter »
(enabled, same link); counts and thumbnail give way to the line.
- Integrity class (R-1): `fingerprint_mismatch`, `drawing_hash_mismatch`, `drawing_missing`, `invalid_document`. Text
A10 in bold, « Contrôle d'intégrité échoué : cette version ne peut pas être affichée. » (EN « Integrity check
failed: this version cannot be displayed. », key `catalogue.card.integrity_failed`), error tone, glyph `alert`, no
« Réessayer ». Storage answering 400 or 404 for a drawing is `drawing_missing` (pinned by a unit test); 401, 429,
network and other Storage answers are not integrity.
- Size limits (D1, owner alternative): `caps` shows the existing « Cette version dépasse les limites de taille d'une
fiche. », error tone, glyph `alert`, no « Réessayer », never the integrity text.
- Other failures (R-2, R-4, R-5): warning tone (the amber of board 30's network card, ink text, no new token), the
existing text (`catalogue.failure.<kind>`; `access_denied.title` « Accès non autorisé » for `denied`; the core's
text for `core_unavailable`), glyph `wifi-off` for `network` and `alert` otherwise, « Réessayer » for that card
only. Once the engine stopped for good the core line has no « Réessayer » and takes the error tone, as on the
sheet page. A `core_unavailable` retry calls `coreRuntime().retry()` before reloading; a version that traps the
core spends one of the tab's three recoveries per click, as on the sheet page. Two calls in the same tick start
one attempt (a ref set by the first call, cleared when its attempt settles).
- Main view not drawable (D2): a version that passes but whose main view the core refuses to compose keeps its
counts and shows the existing `catalogue.failure.view` in a neutral `info` line (no role, no retry). The engine
failing there (`core_unavailable`, `internal`) is the core line with its retry. Nothing is silent any more.
- Announcements and focus (R-3): the lines are not live regions (no burst when several cards fail at load). During
a retry the line and « Réessayer » stay mounted, the button `aria-disabled` and greyed with a wait cursor
(`aria-disabled:opacity-60 aria-disabled:cursor-wait`, no new wording), the card `aria-busy`, clicks ignored.
A retry that fails again keeps the focus on « Réessayer » and a visually hidden polite region, empty at load and
cleared at each attempt, says the line again. The region stays mounted once a retry ran, even when the new
failure offers no retry (network, then `fingerprint_mismatch` or `caps`; the core, then stopped for good), so the
line it ends on is said. When an attempt settles without its « Réessayer » (success, or a failure with no retry
left), the focus moves to the card's « Consulter » only if it was on that card's removed « Réessayer » or on
`body`.
- `Notice` gains the tone `warning`, `icon?` and `live?` (default true); its existing call sites keep their role
and glyph (test). `Icon` gains `wifi-off`. `Card` passes an optional `aria-busy` (outside the plan's file list,
needed by R-3, accepted at review).
- Browser proof (`apps/app/e2e/shell/catalogue-integrity.spec.ts`, `BROWSER_TESTS.md`): one sound and one altered
sheet seeded per run (`storedFingerprint: '0'.repeat(64)`), integrity in fr and en at desktop and 375 px, network
in fr on both projects (one version aborted, `route.fallback()` for the rest, the card remounted by search, never a
reload; both buttons on one row). Cleanup (D4, ruling below): each own sheet's name read by its captured id and
compared exactly with the name the run seeded (« R4 <tag> saine » / « … altérée »; a different name archives
nothing and fails the run), then the sheet archived by id through `private.set_sheet_status` as the seed owner and
read back; the own `catalog_reader` grant revoked by id and read back; no statement matches by pattern. Every id
written into a statement, `seedGrant`'s account id included, goes through `capturedId` (`e2e/support/records.ts`).
- D4 ruling: archive, not soft delete. The owner authorised a soft deletion, but the application never deletes a
sheet that had a published version (§6.9, `delete_anatomical_sheet`), so a deleted row with published history is
a state no screen can reach. Archiving is reachable, takes the sheet out of the catalogue and stays inside the
authorisation. Cost: the archived test sheets stay in the admin sheet list under « Archivée », tagged « R4 <tag> ».
- Rust by default (R-8): sorting closed kinds into lines is display formatting; one class C row in
`RUST_ARCHITECTURE_REVIEW.md` Part II.
- Board 30 deviation found at the design review, fixed in the final fix wave: on desktop (1280 px, three columns,
a card about 288 px wide) the network card stacked « Réessayer » above « Consulter » where the board draws them
side by side. The failure row now narrows the padding of its buttons (`*:px-3`, that row only; « Consulter »
keeps its look elsewhere, the wrap stays as the fallback): at 1280 px « Réessayer » (102 px) and « Consulter »
(125 px) sit on one row, 44 px high, and the network e2e checks it on both projects.
Results (exact; nothing that failed is called green):
| Check | Result |
|---|---|
| App unit suite (`pnpm --filter @faunosteo/app test:run`, at `6ad8d83`) | 1771 tests in 138 files passed (1753 before the card tests) |
| `pnpm typecheck`, `pnpm lint`, `pnpm build` | exit 0 at each code commit (build: the usual chunk-size notice only) |
| `catalogue-integrity.spec.ts` alone on the sandbox (`dadd8e5`) | 6 passed, 0 failed, 0 skipped, 0 flaky, twice (once with temporary screenshot lines, once as committed) |
| `catalogue.spec.ts` alone after the support change | 8 passed, 0 failed, 0 skipped, 0 flaky |
| Data guard before and after the runs (read-only) | 0 removed, 0 changed, 81 added (27 sheets, 43 versions, 11 grants). The 8 « R4 … » sheets read back archived (one `anatomical_sheet.archived` audit row each), the 4 own grants revoked (`e2e cleanup`) |
| Final fix wave: app unit suite (at `3bea19f`) | 1775 tests in 138 files passed (+4: retry ending with no retry left, two calls in one tick, engine stopped for good, two cards on the core) |
| Final fix wave: `pnpm typecheck`, `pnpm lint`, `pnpm build` | exit 0 (build: the usual chunk-size notice only) |
| Final fix wave: `catalogue-integrity.spec.ts` alone | 6 passed, 0 failed, 0 skipped, 0 flaky, twice (once with temporary screenshot lines, once as committed at `3bea19f`) |
| Final fix wave: `catalogue.spec.ts` alone (`seedGrant` changed) | 8 passed, 0 failed, 0 skipped, 0 flaky |
| Final fix wave: data guard (read-only, before and after its three runs) | 0 removed, 0 changed, 81 added (27 sheets, 43 versions, 11 grants), all the runs' own: 8 « R4 … » sheets read back archived (name checked before each archive), 4 own grants revoked (`e2e cleanup`), the rest from `catalogue.spec.ts` and the warm-up as before |
| Catalogue census of the sandbox (read-only, 2026-11-17, below) | 164 entries: 160 pass and compose their main view, 4 `fingerprint_mismatch` (anomaly A) |
| Verifier gap: two tests added (`83f2d8b`: the glyph each line draws; focus after a retry settling on `body` or elsewhere) | app unit suite 1780 tests in 138 files passed; typecheck, lint exit 0 |
| `pnpm check:all` at `2c9d837` (2026-11-17) | exit 0, « all checks passed »: core-wasm 66, supabase 182, app 1780, e2e:ui 253 passed, app image, worker test and runtime image; e2e:shell not run by check:all (the spec runs above are the browser proof) |
Sandbox catalogue census (P1, read-only): every catalogue entry (active official sheet whose current version is
published) read by SQL under `default_transaction_read_only`, its drawings fetched from Storage by GET, then the
core's `PreviewBuilder` built from the version's status, document and stored fingerprint as `loadVersion` does, and
the main view composed as the card does. 164 entries: 80 « Cheval publié <tag> » and 80 « Cheval refusé <tag> »
(left by `catalogue.spec.ts`) pass and compose; the four « Cheval sandbox » entries (two at version 1, two at
version 2) answer `fingerprint_mismatch`: their stored fingerprint is `'a'.repeat(64)`, the fictitious value of the
worker stand-ins (anomaly A, section 5). No other kind was found. Limit: Storage was read with the service key, so a
drawing the reader's Storage policy would refuse is not detected by the census.
Limits (accepted, recorded):
- R-7: when the engine fails for the whole tab, every card shows the core line with its own « Réessayer »; one card's
retry restarts the engine and reloads that card only, the others reload on their own retry.
- R-6: a failed `listSheetVersions` drops the publication date without a word. A failed card shows « Version n »
without a date.
- D2: the view line reuses `catalogue.failure.view`, a text written for the sheet page (« Cette vue ne peut pas être
affichée : … »); no card-specific wording was asked for.
- Sandbox accumulation: the seed owner's row in `private.catalog_write_budgets` is R4's own, not pre-existing
data: there was none before R4, its first run created it, and every later run updates that same row (4 archives
per run, counted per hour and per day; 8 of 300 per hour and 8 of 2,000 per day after the final fix wave's two
runs, outside the data guard); the spec's reader accounts stay (as
every suite's, BLK-015); `catalogue.spec.ts` still leaves its published sheets listed and the global warm-up its
« Fiche e2e … » sheet and `sheet_editor` grant per invocation (existing accumulation, BLK-015, RR-13).
- Every card loads its version at mount: with 164 entries the sandbox catalogue is slow to settle (with RR-13).
- Unit only (rendered cards on the fake client, not in a browser): the engine stopped for good in a card (error
tone, no « Réessayer »); two cards failing on the core (one retry restarts the engine once and reloads that card
only, the other keeps its line and its « Réessayer »); a retry ending on a failure with no retry left (focus on
« Consulter », the polite region says the new line); two retry calls in one tick (one attempt); no blob URL
created after the card left during an attempt. Not checked in a browser: `aria-busy`, `aria-disabled` and the
greyed button during an attempt, a screen reader reading the polite region.
- The cleanup's name check (D4 identity) is proved by the archives that passed (each row's name matched); the
refusal path (a different name archives nothing) is not exercised by a test.
Owner validation 2026-11-17: R4 validated officially after the owner's manual trial on `pnpm db:sandbox trial`
(OWNER_TRIAL §4.16); every step works and the result suits the owner. R4 is closed. The owner confirmed the two
points left open: the spec's test sheets are archived, not deleted (D4 ruling above), and the business rules on
sheets that had a published version stay unchanged; `Card.tsx` (optional `aria-busy`) and `e2e/support/records.ts`
(`capturedId`) are accepted outside the plan's file list. The results above are the final ones. BLK-015, RR-13 and
anomaly A stay open in their own tracking.
## 4. Order and dependencies
R0 first (validation). R1 is independent of R2 and can ship first. R2 depends on the validated boards 23 to 27. R3
ships search and name sort first; its date part waits for the SQL checkpoint (A7 ratified). Superseded 2026-11-16
(D2): R3 shipped in one lot, after M9 was applied to the sandbox and the trial stack.
R4 is independent and may run any time after its board (30) is validated. Anomaly B (section 5) is fixed: the admin specs no longer
depend on the working-copy slot.
## 5. Test anomalies tracked apart (D9)
| Item | Cause | Constraint |
|---|---|---|
| A. Fictitious fingerprints | Worker stand-ins publish with `'a'.repeat(64)`: `packages/supabase/tests/sandbox/sheet-fixtures.ts:115` (`openRound`), `sheets-concurrency.sandbox.ts:168`, `sheets-perf.sandbox.ts:169`. Seven published or deprecated versions on the sandbox fail `check_published`; the owner opened three of them on 2026-11-09. | Compute the real fingerprint through the core as the catalogue e2e does (`fingerprintOf`), never weaken the publish gate or the reader's check. |
| B. Working copy never archived | (a) `packages/supabase/tests/sandbox/sheets-admin.sandbox.ts` created the historical sheet's working copy (`8c315a12-…`, 2026-11-08 21:00:30) and never archived it; the single slot then disables « Créer une copie de travail » (`WorkingCopyAction.tsx:88`). (b) Even on a fresh stack `editor-equine.spec.ts` flips the slot while `admin.spec.ts:717` runs in parallel. Only the historical sheet offers the action, so the test depended on shared state. | FIXED (owner, 2026-11-12, option 1): the dialog's axe, layout and keyboard checks run on the real component in `e2e/ui/working-copy.spec.ts`; `admin.spec.ts` keeps the historical page and checks the button state against its reason; the sandbox suite archives only the copies it created (`returning id` checked) and each cleanup step runs independently. Limit: the dialog is no longer checked in its real server context by this accessibility test. Open: the creation branch of the cleanup is not yet run on a fresh stack. `editor-equine.spec.ts` still archives and restores pre-existing copies on every run. |
| C. Silent card failure | Section 3, R4. | Show the failure, never hide or route around it. Built by R4 (2026-11-17, §R4 « As built »), CLOSED 2026-11-17 by the owner's validation: the card says its failure inside itself, « Consulter » kept, the check unchanged. The four sandbox entries it shows failing are anomaly A, still open. |
## 6. Arbitrations
Status as of the brief « R0 — Décisions propriétaire et dernière validation graphique » (2026-11-09, BDR-108) and the
owner's graphic verdict « R0 — Verdict de revue graphique propriétaire » (2026-11-09, BDR-109): every arbitration is settled.
The owner-facing table with consequences, in French: `docs/design/v1-2-6-r0/ARBITRAGES.md`.
| # | Question | Recommendation | Status |
|---|---|---|---|
| A1 | Reorder commands with a search active in document order | Hidden, as checkpoint choice 10 already rules (moving among a hidden subset is ambiguous); one hint line. | RATIFIED 2026-11-09 (BDR-109), exact recommendation |
| A2 | Tiles open at once; how long open state, search and sort last | Several open at once; kept while the editor stays open (kept-mounted tab), reset when it reopens. | RATIFIED 2026-11-09 (BDR-109), exact recommendation |
| A3 | Actions in the contextual panel | « Modifier », « Retirer de ce constat », « Aller à »; no add, no reorder, no « Supprimer » (always refused by the core there). « Tout afficher » opens the tab with search and sort reset, focus on its heading. | RATIFIED: « Retirer de ce constat » removes only the finding's citation of the reference, never the reference; undoable through the editor history. Since 2026-11-14 the removal is made from the finding's edit form (« Retirer » on the draft list), not from the sheet |
| A4 | Closed tile content | Citation on two lines at most, type, « Citée N fois » or « Non citée » (the count includes families). | RATIFIED 2026-11-09 (BDR-109), exact recommendation |
| A5 | Escape priority | Dialog, then open menu, then contextual panel, then tracing in progress, then selection. | RATIFIED 2026-11-09 (BDR-109), exact recommendation |
| A5b | « Toutes les annotations » already clears the selection | Keep both (D6). | RATIFIED 2026-11-09 (BDR-109), exact recommendation |
| A6 | Focus after « Désélectionner » | The annotation's row in the list. | RATIFIED with precision: the row if it still exists, else the drawing; focus never selects again; an accessible announcement |
| A7 | Source of the last-modification date | SQL read change `last_modified_at`, own checkpoint. | RATIFIED: dedicated SQL checkpoint `V1-2-6_R3_SQL_CHECKPOINT.md`; until then no creation date shown as a modification date and no misleading sort. 2026-11-16: M9 applied (sandbox, trial stack), the two-step delivery superseded by D2 (§R3 « As built ») |
| A8 | Date format | Absolute, « 9 nov. 2026, 12:06 » (EN « 9 Nov 2026, 12:06 »). | RATIFIED 2026-11-09 (BDR-109), exact recommendation; shown only after the SQL checkpoint |
| A9 | Administration home card | The five most recently modified sheets with species, status and date, then « Voir les N fiches ». | RATIFIED 2026-11-09 (BDR-109), exact recommendation; final state after the SQL checkpoint |
| A10 | Catalogue warning wording | FR « Contrôle d'intégrité échoué : cette version ne peut pas être affichée. », a distinct line with « Réessayer » for a load that failed for another reason. | R4 ACCEPTED; wording RATIFIED 2026-11-09 (BDR-109), exact recommendation |
| A11 | Trial stack command | `pnpm db:trial`, prefix `faunosteo_trial_`, ports 58xxx and 5194. | RATIFIED 2026-11-09 (BDR-109), exact recommendation (name, prefix, ports) |
| A12 | Old `db:sandbox trial*` commands | Remove once the trial stack is validated. | DECIDED otherwise: kept for now, neither removed nor silently redirected |
| A13 | Trial accounts | `admin` with catalogue rights, `owner` if needed. | DECIDED 2026-11-09: two accounts; practitioner = technical `admin` + `sheet_editor`, no owner, reviewer or publisher; not a production practitioner profile; O1 unchanged |
| A14 | Key lifetime and secret renewal | One year, `rotate-keys`. | DECIDED 2026-11-09: one-year keys on the dedicated local stack, documented rotation, no secret in Git or logs, security review of the renewal |
| A15 | Backup before every `migrate` | Automatic, kept, never pruned without the owner. | RATIFIED 2026-11-09 (BDR-109), exact recommendation |