Merge chore/step-06-pdf-archive-analysis into develop
ci / check (push) Canceled after 0s
ci / docker (push) Canceled after 0s
ci / supabase (push) Canceled after 0s

This commit is contained in:
bchanot committed 2026-10-11 11:52:14 +02:00
commit 72f19521d3
3 files changed
+632

No files matched your search

+4
View File
@@ -195,3 +195,7 @@ Step 4 validated + finished into develop. User decisions on the 7 open points
- Step 5 addendum (2026-10-10, user testing manually): per-leaf migration tests (612 leaves, anatomy + app + PDF roundtrip), `pnpm report:compare` archive check tool (privacy default, escaping), stale `scaffold:species` script removed, attachment read/parse split. Gates: verifier 2→1→CONFORME, security PASS ×2 (2 MEDIUM + 4 LOW fixed by executors). EVAL-012, LRN-044. Branch 14 commits ahead, not merged; awaiting user validation; no Step 6.
- Step 5 user validation (2026-10-10): all paths OK. One defect: findings picker text blurry on screen. Root cause (headless Chromium measure): permanent `will-change: transform, opacity` on the picker list + fractional offset inside height-animated card → composited text resampled; PDF unaffected. Hotfix 1f0b710: prop dropped, regression test (app 182). Security PASS, docs NONE. Full check green. User signal → finish Step 5; carpus + moved points noted, 28 annotations NOT a final catalog (enrichment with the osteopath + future editor later). Step 6 to present, not started.
## 2026-10-10 (session — Step 5 closed, Step 6 analysis)
Step 5 validated + closed by user; LRN-045 (permanent will-change blurs text) written on develop. Step 6 analysis on chore/step-06-pdf-archive-analysis: 3 analyzers (PDF pipeline, data model, app/tests) → facts: export reads live SheetIndex (labels, views, asset, all annotations of view), 4 fetched logos + 3 hard-coded URLs, DOM in drawing/highlight/raster, attachment = bare v3 Flate'd by pdf-lib, getAttachments dead on pdf-lib 1.17.1, no report↔sheet check anywhere. Measured: lean snapshot envelope 172 KB raw / 51 KB stored (3 obs), 228 / 54 worst; SVG = 90% of stored. Plan: snapshot = lean AnatomicalSheetDocument + assets inline (reuses index/renderer/integrity rules), packages/pdf engine (xmldom compose promoted from app-review.ts, Rasterizer + Branding inputs, no DOM lib), envelope v1 layered reading (report never lost), rebase on import/restore keeps report.sheet == snapshot.ref, export-side caps. 3 plan-challengers GO WITH FIXES (16+19+13 findings folded in). docs/development/STEP6_PDF_ARCHIVE_ANALYSIS.md; TODO 6.0–6.7. Awaiting user decisions D1/D4/D6/D7/D8/D10/D12/D13; no implementation.
+9
View File
@@ -100,6 +100,15 @@ feat/findings-tree-v2 merged (FindingsTreePicker tracked in HEAD); all work belo
- [x] 5.10 final controls on the last lot: stale root script scaffold:species removed; verifier ×3 (2→1→CONFORME) + security ×2 PASS on criteria 16-18; EVAL-012
- [x] 5.11 user validation defect: findings picker text blurry on screen → permanent `will-change` on the list dropped (hotfix 1f0b710, regression test); PDF never affected
- [ ] 5.12 follow-up (not Step 5): equine sheet anatomical enrichment with the osteopath + future sheet editor (Step 17); the 28 draft annotations (carpus added, points moved) are NOT a final catalog
- [ ] Step 6 — autonomous PDF archives (analysis + plan 2026-10-10 on chore/step-06-pdf-archive-analysis: docs/development/STEP6_PDF_ARCHIVE_ANALYSIS.md, challenged on 3 lenses GO WITH FIXES; AWAITING USER VALIDATION of D1/D4/D6/D7/D8/D10/D12/D13 before implementation on feature/step-06-pdf-archives)
- [x] 6.0 analysis: 3 analyzers (pipeline, model, app/tests) + size measurements (envelope 172 KB raw / 51 KB stored for 3 obs, 228 / 54 worst case) + 3 plan-challengers → plan rewritten
- [ ] 6.1 Phase 0 freeze with current code: demo report fixture, migrated-v2-with-unresolved builder, G3 bare v3 PDF, G1 text goldens (logos absent + present), G2 structural facts/placements/links, T12 HomeRoute + T13 EditorRoute characterisation
- [ ] 6.2 Phase 1 domain: sheet-index moved, archive/snapshot (buildSheetSnapshot, assertSnapshot, caps), archive/envelope (schema v1, layered readArchiveDocument, SnapshotRejection), report/rebase (rebaseReportOnSheet), bundle carries asset sha + status/name, tests
- [ ] 6.3 Phase 2 anatomy: render/compose.ts promoted from curation/app-review.ts, highlight on the layer tree, review drawing golden unchanged
- [ ] 6.4 Phase 3 packages/pdf: scaffold (no DOM lib, restricted globals, vitest node, Dockerfile, CI, README), svg/sanitize allowlist + horse.svg zero-removal, render moved, Branding/Rasterizer/exportedAt/safeText/dict, exportReportArchive (T3 both logo modes), readReportArchive (filename-first, layered, attachment_stripped; T4-T6, T8), T1/T7/T9/T10 resvg/T11, xmldom bundle size ≤ 50 KB gz
- [ ] 6.5 Phase 4 app: browserRasterizer, branding.ts, export/import on the engine, rebase on import + draft restore, 2 new messages, old render/attachment/highlight code + pdf.* keys removed, report:compare on engine API, T2/T14/T15, suite migrated 1:1
- [ ] 6.6 Phase 5 docs + gates: PDF-FORMAT.md real, CLAUDE.md map, WORKSPACE.md rule, DEPLOY.md rollback note, CHANGELOG, BDR-014 note; pnpm check, docker build, verifier + security gates, EVAL; manual validation list
- [ ] 6.7 follow-ups recorded for Step 13: catalog asset cap aligned with the archive cap (2 MiB), allowlist sanitiser at upload, app canvas sanitiser (EVAL-011)
- [ ] 5 follow-up: allowlist SVG sanitiser before Storage-served assets (Step 13); orphan findings UX (D5) at the report UX step; report:compare residuals (ZodError reason genericised, v3 lines computed once, long template lines); 16 MiB attachment cap without direct test
---
@@ -0,0 +1,619 @@
# Step 6 — Autonomous PDF archives: analysis and plan
Status 2026-10-10: analysis and plan, challenged on three lenses (correctness,
robustness, simplicity: GO WITH FIXES each, fixes folded in below), awaiting
user validation. No implementation.
Scope authorised by the user on 2026-10-10: the PDF import system, the
attachment and the modules around them may change (this lifts the CLAUDE.md
"ask before" gate for this step). Goal: a Faunosteo PDF must stay readable and
re-renderable when the anatomical sheet it was built on no longer exists in
the catalog. Business behaviour of the PDF is preserved; no graphic redesign.
Sources: three read-only analyses of the current code (PDF pipeline, data
model, app wiring and tests), size measurements on the equine draft, three
plan challenges. Facts carry a `path:line` reference; measurements were taken
on `develop @ 072d761`.
---
## 1. What the current pipeline depends on
### 1.1 Inputs read at export time
`exportReportToPdf(report, sheet: SheetIndex, { locale })`
(`apps/app/src/pdf/export.ts:27`) reads two sources.
From the **live sheet**, through the `SheetIndex` built from the bundled draft:
| Data | Used for | Where |
|---|---|---|
| `annotation.name {fr,en}` (trigger: the opened view's name) | zone text | `render/observationPage.ts:355-356`, `services/sheetIndex.ts:108-111` |
| `finding_zones[].name`, looked up by id whatever the `unresolved` flag | finding group heading | `observationPage.ts:383-390` |
| `view.kind`, `view.name` | "general" label, halo, upper-cased detail caption | `:329-340` |
| `view.view_box`, `view.svg_asset_id`, parent chain | crop and asset choice | `anatomy/drawing.ts:61-85` |
| asset SVG text | the drawing | `drawing.ts:52` |
| `document.views` (names title the nameless triggers), `categories[].color`, every annotation of the view | marker layer, then highlight | `packages/anatomy/src/render/layer.ts:114-255` |
From the **report**: identities, dates, zones `(annotationId, viewId, side,
vertebrae)`, finding leaf snapshots (always the text source, BDR-010) and the
vet flag `(!)` (`observationPage.ts:413-414`), free text, severity,
conclusion. Zone snapshots are a fallback only (`:357-361`).
From the **i18n dict**: `pdf.*`, `severity.*`, `lateral.*`,
`footer.powered_by`, `footer.in_partnership`.
Never read: the finding tree nodes (`nodeAt`), `vertebral_config`,
`report.sheet`, `finding.path`.
### 1.2 Browser and network dependencies
| Dependency | Where | Behind an interface |
|---|---|---|
| `fetch` of 4 logos on root-relative URLs | `render/logo.ts:6`, `export.ts:63` | no |
| `DOMParser`, `document.createElementNS/importNode` | `drawing.ts:37,78,88,95` | no |
| `querySelectorAll`, `el.remove()`, `XMLSerializer` | `observationPage.ts:159-229` | no |
| `DOMParser`, `Blob`, `URL.createObjectURL`, `Image`, canvas, `toBlob`/`toDataURL`, `atob` | `svgToPng.ts:17-101` | no |
| `DecompressionStream` (read side, guarded) | `attachment.ts:127` | no |
| `new Date()` (non-deterministic bytes) | `export.ts:38`, `attachment.ts:35` | no |
Hard-coded branding: 4 logo paths under `public/`, link URLs at
`header.ts:63` and `footer.ts:53,92,107,116`, attachment description,
highlight colour `#b85a3f`. No `useConfig`/`appConfig` on the PDF path.
Logos degrade one by one (own try/catch, `export.ts:47-68`); a single view
raster failure sinks the whole export (`Promise.all`, `observationPage.ts:334`).
Node 24 provides `CompressionStream`, `DecompressionStream`, `TextEncoder`,
`crypto.subtle`, `Blob`, `structuredClone`; it has no `DOMParser`,
`XMLSerializer`, `Image` or `document`. pdf-lib 1.17.1 is DOM-free and already
runs in Node for `pnpm report:compare`. `@faunosteo/anatomy/render` is DOM-free
(`LayerElement` tree + `serializeLayer`), and the curation pipeline already
composes a view drawing with `@xmldom/xmldom` in Node
(`renderAppReviewDrawing`, `packages/anatomy/src/curation/app-review.ts:20-36`:
parse asset, set `viewBox`, drop width/height, append the layer, serialise).
### 1.3 The attachment today
- One embedded file `osteo-report.json`, the bare Report v3 (pretty JSON),
Flate-compressed by pdf-lib itself (`FileEmbedder` → `context.flateStream`,
pako). Full fixture: 3,947 B raw, 1,177 B stored.
- Carries `sheet {sheetId, versionId|null, documentSha256}` and nothing else of
the sheet. No locale, no labels beyond the zone and leaf snapshots.
- Read path: `importReportFromPdf` (50 MiB cap) → `readAttachedDocument`
(16 MiB cap, raw object traversal: the `getAttachments` branch never runs on
pdf-lib 1.17.1, so the filename is never checked and the first
`EmbeddedFile` wins) → `readReportDocument` (v3 parsed, v2 migrated with the
bundled map, else `incompatible_version`). Any other failure becomes
`no_attachment` ("not generated by Faunosteo").
- Nothing compares `report.sheet` with the loaded sheet, on import or on
export. Every `pnpm import:equine` run changes the draft's `documentSha256`.
### 1.4 What the report already preserves
Report v3 (`packages/domain/src/report/v3.ts`) embeds: zone names FR/EN
(optional), leaf finding names FR/EN (required), the vet flag, vertebra codes
and indexes, the side, `unresolved: true` on items the v2 migration could not
place (their legacy ids are kept; the 29 unresolved legacy leaves all sit in
zones that still exist: frontal, sphenoid, ssb). It does **not** embed
finding-zone names, ancestor and category names, region names, view names,
colours, geometry, view boxes or assets.
---
## 2. Snapshot: exact content
### 2.1 Requirement
Re-read and re-render an old bilan (FR and EN, text and drawings) without the
current catalog, and keep every historical datum including unresolved items.
A sheet id or a dependency on the current catalog is not enough.
### 2.2 Shape: a lean sheet document, not a new schema
The snapshot has the shape the renderer and the sheet index already consume
(`LoadedSheet` minus the migration map): `{ status, ref, document, assets }`.
`document` is a valid `AnatomicalSheetDocument` (schema v2, strict, unchanged)
with the bibliography emptied and the finding trees pruned to what the report
uses. So `renderAnnotationLayer`, `buildSheetIndex`, `pickLabel` and
`validateSheetDocument` apply to it with no change, and the full set of sheet
integrity rules (unique ids, acyclic view chain, one main view, categories and
regions present) protects it.
| Part | Content | Why |
|---|---|---|
| `ref` | `sheet_id`, `version_id`, `document_sha256`, `schema_version`, `species_id`, `name {fr,en}`, `status` | provenance (`document_sha256` is the full catalog document's hash, kept as a pointer, not as a hash of the snapshot) |
| `document.views` | all views | captions, crops, trigger titles (a nameless trigger is titled from the view it opens, `layer.ts:240-248`) |
| `document.annotations` | all annotations, `review_status` kept (schema default) | geometry, colours, zone names, anatomical relations (`finding_zone_id`, `vertebral_region_code`, `opens_view_id`, `laterality`) |
| `document.categories`, `vertebral_config` | all | marker colours, region names and counts |
| `document.finding_zones` | every zone `{id, name}`; for a zone referenced by any finding of the report, resolved or `unresolved`, its categories pruned to the referenced paths (`{id, kind, name}` and every node on the path `{id, name, requires_vet_referral}`, `description: {}`, `reference_ids: []`); other zones keep `categories: []` | group headings by id; the labels of every ancestor on a used path are preserved (+4 KB raw, +0.4 KB stored over empty trees) |
| `document.references` | `[]` | bibliography is not bilan data (15 KB) |
| `document.name`, `description`, `scope_disclaimer`, `attribution` | kept as is (1.4 KB) | sheet identity and legal scope of the historical sheet |
| `assets` | `[{ svg_asset_id, sha256, svg }]` for every asset reachable from a view | the drawing; `sha256` is the catalog's content address copied from the seed manifest (`assets.json`), carried by the bundle from now on, a provenance pointer and not a tamper check |
The SVG text goes **inline as a JSON string** (SVG is UTF-8 text: no base64
penalty, one attachment, one size cap, one parse). Separate embedded files
were rejected: the raw-object reader takes the first `EmbeddedFile`, and
several attachments multiply the stripping risk.
Rejected alternative: a dedicated `SheetSnapshotSchema`. It would force a
rewrite of every sheet lookup in the render path, leave `renderAnnotationLayer`
without a valid input, and lose the sheet integrity rules.
### 2.3 Envelope (attachment v1)
```json
{
"format": "faunosteo-report-archive",
"envelope_version": 1,
"exported_at": "2026-10-10T18:00:00.000Z",
"locale": "fr",
"report": { "version": 3, "...": "Report v3 unchanged" },
"sheet_snapshot": {
"status": "draft",
"ref": { "sheet_id": "…", "version_id": null, "document_sha256": "…", "schema_version": 2, "species_id": "equine", "name": { "fr": "…", "en": "…" } },
"document": { "schema_version": 2, "...": "lean AnatomicalSheetDocument" },
"assets": [ { "svg_asset_id": "…", "sha256": "…", "svg": "<svg …>" } ]
}
}
```
Report v3 is wrapped, not changed: no report version bump. Invariant written
and checked: `report.sheet.documentSha256 === sheet_snapshot.ref.document_sha256`
(see §3.5 for how the app keeps it true). Filename stays `osteo-report.json`
(CLAUDE.md); description becomes `Faunosteo report archive v1`; the PDF Info
dictionary gains `Keywords: faunosteo-archive/1` so a PDF whose attachment was
stripped by a viewer can still be recognised. Pretty-printed as today (an
extracted archive stays human-readable). `locale` and `exported_at` make the
attachment differ between locale variants: BDR-014's "byte-identical JSON"
note is superseded on that point, the report inside stays identical.
### 2.4 Measured sizes (equine draft, pretty JSON, deflate 6 ≈ pdf-lib Flate)
| Envelope | Raw | Stored |
|---|---|---|
| 3-observation fixture, lean document + SVG | 172 KB | 51 KB |
| Worst case 27 zones × 3 findings, both views | 228 KB | 54 KB |
| Lean document alone, 3 observations | 29 KB | 4 KB |
| Full document (bibliography included) + SVG, alternative | 464 KB | 79 KB |
| SVG alone, as a JSON string | 138 KB | 47 KB |
The drawing is 90% of the stored size. The 250 KB target holds on the raw
size (worst case 228 KB) and with a 4× margin on the stored size. Data
conservation wins over size: nothing in §2.2 is dropped to fit.
---
## 3. Target architecture
### 3.1 One sheet input for the engine: the snapshot
```
app export : LoadedSheet + Report ──buildSheetSnapshot()──▶ SheetSnapshot ─┐
archive re-export: envelope.sheet_snapshot ─────────────────────────────────────┤
▼
exportReportArchive({ report, sheet, branding, locale, rasterizer, exportedAt }) ─▶ { pdf, warnings }
```
The engine builds its `SheetIndex` from the snapshot and never sees a live
sheet. The app builds the snapshot from the loaded sheet at export; an archive
re-export passes `envelope.sheet_snapshot` as is. The attachment embeds exactly
what the engine consumed.
Label policy, identical to today's output:
- zone name = snapshot annotation name by id (trigger: the opened view's
name), else the report's zone snapshot in the requested locale, else the id;
`unresolved` is never consulted;
- finding group heading = snapshot finding-zone name by id, else the id
upper-cased (as today);
- finding text and the `(!)` marker always come from the report (BDR-010);
- views: unique view ids of the observation present in the snapshot, main
first; an unknown view prints no image, as today.
### 3.2 Packages and moved code
| Package | Content |
|---|---|
| `packages/domain` | `archive/envelope.ts`: `ReportArchiveEnvelopeSchema` (v1, metadata tolerant to unknown keys, `report` and `sheet_snapshot` at fixed keys), snapshot caps, `readArchiveDocument(data, map)` with layered reading (§4), `ArchiveReadError`; `archive/snapshot.ts`: `buildSheetSnapshot(loadedSheet, report)` (pure: prune trees, empty references, pick assets), `assertSnapshot(snapshot)` (schema, caps, `validateSheetDocument` with no error, every view reaches an asset present in `assets`); `anatomical-sheet/sheet-index.ts`: `buildSheetIndex`, `pickLabel`, `isObservable` moved from `apps/app/src/services/sheetIndex.ts` (the app re-exports them); `report/rebase.ts`: `rebaseReportOnSheet(report, index, archivedSnapshot?)` (§3.5) |
| `packages/anatomy` | `render/compose.ts`: `composeViewSvg({ assetRoot, document, viewId, highlight })` promoted from `curation/app-review.ts` (xmldom), highlight applied on the `LayerElement` tree before it becomes DOM (fill `#b85a3f`, opacity 0.85, `osteo-halo` filter on the main view, unselected markers dropped, no `data-*`, `<title>` kept); the curation review drawing calls it with no highlight |
| `packages/pdf` (new, `@faunosteo/pdf`) | the engine: `exportReportArchive`, `readReportArchive`; `render/*` moved as is (metadata, fonts, pages, header, footer, text, link annotations, constants); `svg/sanitize.ts` (allowlist, xmldom, §6); `svg/view.ts` (parse → sanitise → compose → serialise, raster size from the view box); `attachment/{write,read,inflate}.ts`; `text.ts` (`safeText`, §6); `dates.ts` (`formatDate` copy, never throws on a free string); `i18n/{fr,en}.ts` (verbatim copies of today's `pdf.*`, `severity.*`, `lateral.*`, `footer.powered_by`, `footer.in_partnership`); types `Branding`, `Rasterizer`, `ExportInput`, `ExportResult`. `tsconfig`: `lib: ["ES2022"]`, `types: ["node"]`; ESLint on `src/**`: `no-restricted-globals` (`fetch`, `window`, `document`, `Image`, `DOMParser`, `XMLSerializer`, `navigator`, `Buffer`, `process`, `require`) and `no-restricted-imports` (`@/*`, `react`, `node:*`); vitest `environment: node` |
| `apps/app/src/pdf` (shrinks) | `export.ts` (snapshot + branding + browser rasterizer → engine; download unchanged), `import.ts` (engine read → `ImportError` kinds, BDR-015), `browserRasterizer.ts` (today's `svgToPng`), `branding.ts` |
| `apps/app/scripts` | `report:compare` moved onto the engine read API (unchanged options) |
Dependencies added:
- `@xmldom/xmldom` 0.9.x as a runtime dependency of `packages/pdf` and
`packages/anatomy` (`compose.ts`), pinned: parse, sanitise, compose and
serialise without a browser, in the browser too (one code path). Already in
the workspace as a devDependency of `packages/anatomy` (0.9.12). Unminified
276 KB / 71 KB gzipped, 43 KB of which is the HTML entity table; expected
35–45 KB gzipped in the app bundle, measured in Phase 3 with a stated
ceiling of 50 KB, else the browser keeps the native DOM for composition.
Parsed with an `onError` that throws on `error` and `fatalError` and never
logs raw input. Watched by `pnpm audit --prod`.
- `@resvg/resvg-js` as a **devDependency** of `packages/pdf` (D6): real
rasterisation in Node for one end-to-end test. Native prebuilt binaries
(lockfile must carry `linux-x64-musl` for the Docker builder stage, where
the unfiltered install pulls it; the runtime image is nginx and unaffected).
No text-extraction dependency: drawn-text goldens reuse the
`PDFPage.prototype.drawText` spy pattern (LRN-033).
### 3.3 Browser-free guarantees
1. Compile-time: `packages/pdf` has no `DOM` lib, so `document`, `Image`,
`DOMParser` do not type-check.
2. Lint-time: restricted globals and imports on `src/**` (tests may read
fixtures with `node:fs`).
3. Test-time: the engine suite runs under vitest `environment: node` with no
Vite sheet plugin, so nothing can reach the bundled sheet.
4. Rasterisation is an adapter:
`type Rasterizer = (svg: string, size: { widthPx: number; heightPx: number }) => Promise<Uint8Array>`.
The engine computes the size (width 1200 px, height from the view box,
aspect clamped to [0.2, 2.0]) and degrades per view: a failed raster gives
a page without that image and a `warnings` entry, never a failed export.
Logos degrade one by one as today.
### 3.4 Branding as explicit input
```ts
interface Credit { logoPng: Uint8Array | null; url: string; fallbackText: string }
interface Branding {
siteUrl: string;
headerLogoPng: Uint8Array | null;
footerLogoPng: Uint8Array | null;
poweredBy: Credit | null; // today: ZenQuality, 70×16 pt
partner: Credit | null; // today: Paws in Harmony, aspect-derived width
}
```
The app assembles it once per export in `branding.ts`: PNG bytes from
bundled assets (`?url` import + fetch inside the app, allowed there), the Paws
SVG rasterised with the browser rasterizer before the engine runs, URLs and
fallback handles as app constants (none of them is an identity string of
`ConfigProvider`; the PDF prints no site name). Branding is **not** embedded in
the archive: brand assets are the publisher's, not bilan data, and a re-export
uses the branding it is given.
### 3.5 Import flow and the sheet invariant
`readReportArchive(pdfBytes, migrationMap)` returns
`{ report, snapshot | null, origin: 'envelope_v1' | 'bare_v3' | 'v2', migrated, unresolvedZones, unresolvedFindings, snapshotRejected?: SnapshotRejection }`
with `SnapshotRejection = 'unsupported_envelope' | 'schema' | 'caps' | 'integrity' | 'sheet_mismatch'`
(a closed enum: no received value is echoed). Attachment selection: the
Filespec whose `/F` starts with `osteo-report` first, then the first
`EmbeddedFile` (today's behaviour as fallback); the dead `getAttachments`
branch is removed. A PDF stamped `faunosteo-archive/1` with no attachment
gives the new `ImportError` kind `attachment_stripped` (message: the viewer or
optimiser removed the data; re-export from the original).
**Invariant and rebase.** The editor always works on the loaded sheet. When a
report arrives whose `report.sheet.documentSha256` differs from the loaded
ref (import of any archive, or a session draft restored after the bundle
changed), the app calls `rebaseReportOnSheet(report, index, archivedSnapshot?)`:
- a zone whose `annotationId` is absent from the loaded sheet, or whose
`viewId` is unknown, gets `unresolved: true`; its label snapshots are filled
from the archived snapshot when present (labels survive the rebase);
- a finding whose `(findingZoneId, path)` has no node in the loaded sheet gets
`unresolved: true` (same semantics as the v2 migration);
- `report.sheet` becomes the loaded ref; counts are returned and the existing
partial-import alert pattern shows them (new key `home.import.rebased`,
wording to validate: "Bilan établi sur une autre version de la fiche.
Éléments non retrouvés : {count} (conservés avec leurs libellés).").
Export then refuses when the invariant does not hold (`ArchiveExportError
'sheet_mismatch'`), which cannot happen through the app once rebase runs on
every entry path. An envelope whose two references disagree is imported with
`snapshotRejected: 'sheet_mismatch'`.
The imported snapshot is not kept in the session (D8): once rebased, the
report lives on the current sheet, the original PDF remains the historical
rendering, and autonomy is exercised by the engine API and the tests.
---
## 4. v2 / v3 archive compatibility
Layered reading, so the report is never lost because of its wrapper:
1. Locate the report: `data.report` when `data.envelope_version` is numeric,
else `data` itself (bare).
2. Parse the report with `readReportDocument` (v3 parsed, v2 migrated with the
bundled map, any other `version` → `incompatible_version` with the report
version found). `foundVersion` always names a **report** version, so the
existing UI message stays true.
3. Parse the envelope metadata tolerantly (unknown keys ignored within v1).
`envelope_version` other than 1 → `snapshotRejected: 'unsupported_envelope'`,
report kept. Then parse `sheet_snapshot` strictly (document schema v2,
caps, integrity, assets reachable, `sheet_mismatch`); any failure →
`snapshotRejected` with its reason, report kept.
| Archive | Produced by | Import | Re-export |
|---|---|---|---|
| v2 bare report | app before Step 5 | migrated with the map; unresolved items flagged, snapshots and legacy ids kept; then rebased (no-op when the map's sheet is the loaded one) | envelope v1 on the current sheet snapshot; unresolved items print from their snapshots |
| v3 bare report | Step 5 develop | parsed, `snapshot: null`, rebased if the draft changed since | as above |
| envelope v1 | Step 6 | report parsed first, snapshot validated or rejected with a reason, rebased if the sheet changed | from the embedded snapshot, no sheet needed |
| envelope v≥2 | future | report imported if its version is known, `snapshotRejected: 'unsupported_envelope'` | — |
| report v≠2,3 | future | `incompatible_version` with that version | — |
Durability rule: archive readers never drop a schema version. Sheet document
schema v2 stays readable in `packages/domain` for as long as envelope v1
archives exist, exactly as Report v2 stays readable today. Forward rule:
additive optional fields keep `envelope_version: 1` (readers ignore unknown
metadata keys); a breaking change bumps it and the reader keeps every older
version. A rollback to a pre-Step-6 app reads an envelope as `no_attachment`;
DEPLOY.md records that limit.
A v2 archive has no legacy sheet snapshot. Its own pages are the historical
rendering; the attachment serves data recovery. Building a "legacy snapshot"
from `legacy-import/` for a faithful re-render of v2 bilans is possible (the
baseline document and the two legacy SVGs are in the repo) but not proposed:
no external users, BDR-038 rejected v3→v2 work, and the printed PDF already
holds the picture. D4 covers keeping the original v2 JSON.
Fixtures: the two frozen v2 PDFs stay; Phase 0 freezes, with the current code
(LRN-043): a **v3 bare golden** PDF, a **demo report** `report-v3.demo.json`
(6 observations: main and head views, vertebrae, bilateral sides, a
vet-referral leaf, a leaf label with an en dash, free text, next appointment)
and a **migrated v2 report with unresolved items** (sphenoid / ssb leaves
reached only through unresolved paths).
---
## 5. Re-export without current resources
`exportReportArchive({ report: envelope.report, sheet: envelope.sheet_snapshot, branding, locale, rasterizer, exportedAt })`
is the whole answer: no sheet provider, no Vite, no browser, no network. The
labels, drawings, colours, crops and relations come from the snapshot; the
finding texts from the report; the only external inputs are the branding and
the locale. Test T1 (§7) performs exactly this call in a Node environment
where no sheet exists. A future server worker runs the same call.
An operator CLI (`report:reexport`) is **not** planned in this step (D12): the
test proves the property, and a CLI is a public command name plus a file
writer over untrusted content (output names must derive from the input file
name, `wx` open, LRN-044 escaping, resvg under memory and time limits).
---
## 6. Size and security limits
Write side (new): `buildSheetSnapshot` output goes through `assertSnapshot`
(schema, caps, integrity, assets present, sanitiser acceptance of every asset,
sheet invariant) and the serialised envelope must stay under the 16 MiB
attachment cap; otherwise the export fails with a typed `ArchiveExportError`
(`snapshot_invalid | snapshot_too_large | sheet_mismatch | too_many_pages`)
and an alert, never a file that cannot be read back. The equine draft is 60×
under every cap. Step 13 must align the catalog's asset cap with the archive
cap (BDR-032 allows 20 MiB assets today).
Caps (checked on both sides, Zod and counters):
| Item | Cap |
|---|---|
| PDF file | 50 MiB (existing) |
| attachment, raw and inflated | 16 MiB (existing, bomb guard) |
| assets | ≤ 4 per snapshot, ≤ 2 MiB each, ≤ 6 MiB total |
| document | ≤ 64 views, ≤ 1,024 annotations, ≤ 8,192 finding nodes, depth ≤ 8 (pre-checked before the recursive schema), labels ≤ 512 chars |
| report at export | ≤ 500 observations, ≤ 8 views per observation (`too_many_pages`) |
| sanitiser | ≤ 20,000 elements, nesting ≤ 64, ≤ 256 `<use>`, filter `stdDeviation` ≤ 50 |
| raster | width 1200 px, aspect clamped to [0.2, 2.0] (an untrusted root `viewBox` cannot request a 120-million-pixel canvas) |
SVG sanitiser (`packages/pdf/src/svg/sanitize.ts`, xmldom, used at export on
the bundled asset and at re-export on the archived one; the app canvas is out
of scope, EVAL-011 keeps it for Step 13):
- the text is refused if it contains `<!DOCTYPE` or `<!ENTITY`;
- element allowlist, enumerated from the real inventory of `horse.svg` plus
structural needs: `svg g path circle ellipse rect line polyline polygon defs
linearGradient radialGradient stop filter feGaussianBlur feMerge feMergeNode
feOffset feFlood feComposite feBlend feColorMatrix clipPath mask title desc
switch use`; no `style`, `script`, `image`, `a`, `foreignObject`, `text`,
`feImage`, `feTurbulence`, `feConvolveMatrix`;
- attribute allowlist from the same inventory: `id`, geometry, presentation
(`fill stroke* opacity fill-opacity fill-rule transform style`), gradient
and filter attributes (`gradientUnits gradientTransform offset stop-color
stop-opacity in in2 result stdDeviation color-interpolation-filters x y
width height`), `viewBox`, `xmlns*`, `systemLanguage` (the `<switch>`),
`font-*`; no `on*`; `href`/`xlink:href` and `url(…)` values must be
`#fragment`; `style` attributes may not contain `url(` other than `#`,
`expression(` or `@import`;
- returns the list of removals; the engine logs nothing from it and the test
on `horse.svg` asserts an empty list and a byte-identical serialisation.
Other guards: strict document schema, `Object.hasOwn` on records, depth
pre-check, no new PII in the envelope (export date, locale). WinAnsi: one text
ingress `safeText()` used by measurement and drawing, testing each code point
with `Encodings.WinAnsi.canEncodeUnicodeCodePoint` (en and em dashes are
encodable and must survive), replacing the rest by `?` and stripping control
characters; a canary test covers `⚠`, `→`, an emoji, `\t`, U+2028 (restores
the lost EVAL-003 test). `formatDate` never throws: a non `YYYY-MM-DD` string
is printed as is.
---
## 7. Tests that prove autonomy
Phase 0 goldens and characterisation tests, frozen with the current code
before any change:
- G1 drawn-text sequence per page, FR and EN, for: the v2 full fixture, the
demo report, the migrated v2 report with unresolved items. Two harness
modes: logos absent (fetch 404, as today) and logos present (fetch stub
returning the real PNG bytes from `public/`, Paws raster stubbed).
- G2 per view, structural: `viewBox`, highlighted ids with
`fill`/`fill-opacity`/`filter`, hidden marker count, halo filter on the main
view only, no `data-*` left; plus image placements `(page, x, y, w, h)` and
link annotations `(page, Rect, URI)`.
- G3 `report-v3.bare.pdf` exported by Step 5 code (import fixture only).
- T12 `HomeRoute` import: v2 alert with count, error kinds to messages,
navigation. T13 `EditorRoute` export: call arguments, download name.
Both written now on the current routes and kept green unchanged.
Engine (`packages/pdf`, node environment):
- T1 **No-sheet re-export**: export the demo report through the app path →
read the envelope → `exportReportArchive` from the envelope alone → same G1
text, same G2 facts, same envelope minus `exported_at` (the written snapshot
is `assertSnapshot`'s parsed output, so key order is stable).
- T3 G1/G2 equality between the new engine and the Phase 0 goldens in both
logo modes: proof of "no visible change".
- T4 roundtrip envelope → import → `toEqual` report and snapshot.
- T5 v2 fixtures import (29 unresolved on the 612-leaf report, zones `[]`
everywhere, so an empty-views observation must export) → re-export →
import again equals; G3 imports with `snapshot: null`.
- T6 versions: envelope v2 with report v3 → report imported,
`snapshotRejected: 'unsupported_envelope'`; envelope v1 with report v4 →
`incompatible_version` 4; bare v7 → `incompatible_version` 7.
- T7 size: demo envelope stored < 250 KB, raw size printed.
- T8 security: 5 assets, a 3 MiB asset, depth 9, 1,025 annotations, a
view chain cycle, a dangling `svg_asset_id`, `sheet_mismatch` → the report
imports and `snapshotRejected` names the reason; hostile SVG (`<script>`,
`onload`, `javascript:` href, external `<image>`, `<!ENTITY>`, `<style>`
with `url(http…)`, 257 `<use>`, a root `viewBox` of `0 0 1 100000`)
stripped, refused or clamped; `__proto__` ids; 24 MB inflate bomb
(existing); export-side: a snapshot over caps fails the export with a typed
error and no file.
- T9 WinAnsi canary (§6) on labels, user text, branding fallback text.
- T10 real raster (resvg) on the demo report: PNG dimensions follow the
1200 px rule and the clamp, highlight colour present, non-white ratio in
range for main and head views; a raster failure yields a page without image
and one warning.
- T11 (CI, not a test): `pnpm typecheck` and `pnpm lint` of `packages/pdf`
pass with no DOM lib and the restricted globals.
App:
- T2 **Catalog changed**: mutate the loaded sheet's labels and one geometry,
re-export the frozen envelope through the app's engine call → G1/G2 of the
envelope, not of the mutated sheet.
- T12, T13 kept green; T15 rebase: import a v3 archive whose `documentSha256`
differs → unresolved flags, labels filled from the archived snapshot,
`report.sheet` updated, alert shown; draft restore path covered too.
- T14 i18n parity: the engine dict FR/EN share one key set; app dicts stay in
parity after losing `pdf.*`.
Manual validation: export the demo report in FR and EN, open in two viewers,
check pages and the attachment panel (`osteo-report.json`), import on a fresh
tab, import a pre-Step-5 v2 PDF, re-import the exported file after
`pnpm import:equine` changed the draft and read the rebase alert.
---
## 8. Visible changes
In the PDF pages: **none** by default; T3 enforces it. Proposed, not planned:
- V1 provenance line in the footer or on the conclusion page: sheet name,
version and status ("Fiche équine, brouillon v1") and export date. Useful
on paper, honest about the draft status. Needs validation.
- V2 marker on items printed from snapshots because the sheet no longer holds
them, for example `(?)` before the label, mirroring the `(!)` vet marker.
Needs validation.
Outside the pages: attachment description "Faunosteo report archive v1" and
the `Keywords` stamp in the PDF properties; in the app, two new messages
(`home.import.rebased`, `home.import.error_attachment_stripped`), wording to
validate.
---
## 9. Decisions to validate
| # | Decision | Recommendation | Alternative |
|---|---|---|---|
| D1 | Finding trees in the snapshot | pruned to the paths the report uses, ancestor labels kept (+4 KB raw) | empty trees (ids and leaf labels only, in the report) |
| D4 | v2 provenance | do not embed the original v2 JSON in re-exports (the original PDF remains the record; the migration is reversible through the committed map) | optional `origin` block with the verbatim v2 document |
| D6 | Node rasterizer | `@resvg/resvg-js` devDependency for T10 | no real raster in Node (structure tests only) |
| D7 | Branding | explicit input, not archived | archive the branding bytes (+90 KB per PDF) |
| D8 | Imported snapshot in the app | not kept; report rebased onto the loaded sheet with the alert of §3.5 | keep it in the session and offer "re-export the archive as is" (new UI) |
| D10 | Visible changes V1, V2 | none in this step | V1 and/or V2 |
| D12 | `report:reexport` CLI | not in this step | add it, envelope v1 only, with the file-writing rules of §5 |
| D13 | New app messages | `home.import.rebased` and `error_attachment_stripped` as worded in §3.5 and §8 | other wording |
Settled in this plan (no observable effect): inline SVG text, lean-document
snapshot shape, xmldom composition everywhere, engine-owned dictionary,
demo report as the DoD reference, `exportedAt` as an input (no clock
abstraction).
---
## 10. Risks
| Risk | Mitigation |
|---|---|
| xmldom serialisation differs from `XMLSerializer` | G2 compares structural facts, not strings; T10 covers pixels; the browser rasterizer re-parses anyway |
| xmldom bundle cost above 50 KB gzipped | measured in Phase 3; fallback: native DOM composition in the browser behind the same `composeViewSvg` signature |
| resvg renders the 81 Gaussian-blur filters or `<switch>` differently from Chromium | T10 asserts structure, not pixels; the browser stays the production rasterizer |
| Moving render code touches 8 modules and 6 test files | one-to-one test migration with the replacing test named (Step 1 rule) |
| Rebase marks items unresolved that only moved on the drawing | by design: ids that survive stay resolved; the alert reports counts only |
| A future sheet with several large assets | per-asset and total caps with an export-time typed error; Step 13 aligns the catalog cap |
| Allowlist strips legitimate content of a future asset silently | the sanitiser returns its removals; Step 13 runs it at upload and fails there (TODO recorded as a `shortcut:`) |
---
## 11. Plan
Executors as in Step 5: one executor per phase with a closed contract, TDD,
fresh verifier and security gates at the end, no publication, no graphic
redesign. Branch `feature/step-06-pdf-archives`.
**Phase 0 — Freeze (current code, nothing else changes)**
- 0.1 fixtures: `report-v3.demo.json`, migrated-v2-with-unresolved report
builder, G3 `report-v3.bare.pdf`
- 0.2 goldens G1 (both logo modes), G2 (structural facts, placements, links);
one-shot `MAKE_GOLDEN=1` run, generator deleted (LRN-043)
- 0.3 characterisation tests T12 (HomeRoute import) and T13 (EditorRoute
export)
**Phase 1 — Domain: snapshot, envelope, rebase**
- 1.1 `anatomical-sheet/sheet-index.ts` moved from the app (re-export kept)
- 1.2 `archive/snapshot.ts`: `buildSheetSnapshot`, `assertSnapshot`, caps
- 1.3 `archive/envelope.ts`: schema v1, `readArchiveDocument` layered
reading, `ArchiveReadError`, `SnapshotRejection`
- 1.4 `report/rebase.ts`: `rebaseReportOnSheet`
- 1.5 bundle carries the asset manifest sha and the sheet `status`/`name`
(`vite/sheet-bundle.ts`, `sheets/provider.ts`, `LoadedSheet`)
- 1.6 tests: pruning on the draft (counts, bytes), `validateSheetDocument`
clean on the lean document, caps, dispatch table of §4, rebase cases
**Phase 2 — Anatomy: composer**
- 2.1 `render/compose.ts` promoted from `curation/app-review.ts`, highlight
on the layer tree; review drawing unchanged (golden test proves it)
**Phase 3 — Engine `packages/pdf`**
- 3.1 scaffold: manifest, tsconfig, lint rules, vitest node, Dockerfile
manifest copy, CI, `packages/README.md` (fix its two wrong lines while
there)
- 3.2 `svg/sanitize.ts` + `svg/view.ts`; `horse.svg` zero-removal test;
hostile cases
- 3.3 render modules moved; `Branding`, `Rasterizer`, `exportedAt`,
`safeText`, dict; parallel rasterisation then sequential embedding
- 3.4 `exportReportArchive` (snapshot-based index, label policy §3.1,
per-view degradation, caps, envelope write, Keywords stamp) → T3 in both
logo modes
- 3.5 `readReportArchive` (filename-first selection, inflate cap, layered
reading, `attachment_stripped`) → T4, T5, T6, T8
- 3.6 T1, T7, T9, T10 (resvg), T11; bundle-size measurement of xmldom
**Phase 4 — App integration**
- 4.1 `browserRasterizer.ts`, `branding.ts`, `export.ts` and `import.ts`
rewritten on the engine; rebase on import and on draft restore; two new
messages
- 4.2 old `src/pdf/render`, `attachment.ts`, `anatomy/drawing.ts` highlight
code and `pdf.*` app keys removed; `report:compare` on the engine read API
- 4.3 T2, T14, T15; T12/T13 green unchanged; app suite migrated one-to-one
**Phase 5 — Docs and gates**
- 5.1 `docs/PDF-FORMAT.md` written for real (pages, envelope v1, snapshot
content, caps, read order, compatibility table, rebase rule, CLI);
CLAUDE.md folder map and conventions; `WORKSPACE.md` package rule corrected
(tests under `tests/`); DEPLOY.md rollback note; CHANGELOG; BDR-014 note
- 5.2 `pnpm check`, docker build, verifier and security gates, EVAL
- 5.3 manual validation list (§7) for the user
Definition of Done: no fetch and no DOM in `packages/pdf` (compile, lint,
tests); T1–T15 green; Phase 0 goldens reproduced in both logo modes; demo
envelope < 250 KB stored and raw; `docs/PDF-FORMAT.md` complete; v2 fixtures
import and re-export; rebase alert shown when the draft changed.