Task 4 report: persist and cache one rich vision read per page

RED evidence

Command:

npx vitest run src/lib/dn-pipeline/extract.test.ts src/lib/dn-pipeline/page-store.test.ts

Result before implementation: failed as expected. The new page-store suite could not resolve ./page-store, and the rich parser assertions received undefined for the new fields (2 failed tests, 1 failed suite).

GREEN evidence

Focused verification:

npx vitest run src/lib/dn-pipeline/extract.test.ts src/lib/dn-pipeline/page-store.test.ts

Result: Test Files 2 passed (2), Tests 11 passed (11).

Regression verification:

npx vitest run src/lib/dn-pipeline/extract.test.ts src/lib/dn-pipeline/page-store.test.ts src/lib/dn-pipeline/run.test.ts src/lib/dn-pipeline/segment.test.ts

Result: Test Files 4 passed (4), Tests 58 passed (58).

npm run typecheck

Result: passed (tsc --noEmit).

git diff --check

Result: passed.

Implementation

Extended PageRead and the extraction schema/prompt with supplier/site text, PO candidates, printed pagination, vehicle registration, raw text, and evidence while retaining all existing commercial fields.
Added defaults for absent rich fields and preserved parser failures as thrown errors.
Added readOrReusePage() with schema/prompt-version cache reuse, reading, successful extraction persistence, bounded failure persistence, and rethrowing of extraction errors. Persisted results are schema-validated before being stored.
The extractor continues to send canonical bytes as data:image/png.

Self-review and concerns

Only Task 4 files plus this report were changed; Clock/photo-app files were untouched.
Rich PageRead fields are required in the TypeScript interface; existing page fixtures/builders now provide explicit null/empty defaults.
Cache reuse trusts a stored extraction when status and schema/prompt versions match, as required; it does not invoke the vision extractor.

Commit

0bd344a80e536d807c75806947e8a343a53ca171 feat(dn): persist cached page vision reads

Fix round 1

Failure-status persistence errors are surfaced with both the extraction and PocketBase persistence messages; they are no longer silently swallowed.
Matching-version cached extractions are validated through parsePageRead(); corrupt cache data is treated as a miss and re-extracted.
Added tests for corrupt cache recovery and failure-status write errors.
Task 7 owns staged-runner integration; Task 4 intentionally does not wire readOrReusePage() into legacy run.ts.

Verification:

npx vitest run src/lib/dn-pipeline/extract.test.ts src/lib/dn-pipeline/page-store.test.ts src/lib/dn-pipeline/run.test.ts src/lib/dn-pipeline/segment.test.ts

Result: Test Files 4 passed (4), Tests 60 passed (60).

npm run typecheck passed (tsc --noEmit); git diff --check passed.

Fix round 1 commit: ea791d6720a1fa21a548c4517e2834790a2141c4 fix(dn): harden cached page read persistence.

Fix round 2

Report-only correction: this round updates commit provenance in this report only. No implementation or test files were changed.

3.1 KiB Raw Permalink Blame History