GSAP carries the counter roll-ups, the bandeja card physics, the dialog
transitions and the three success moments FLOWS.md allows. Every one of
them checks prefers-reduced-motion first and does nothing when it is set.
boneyard and canvas-ui are not what SPEC.md's stack table says they are:
on npm the names belong to two abandoned projects that do neither job.
The skeletons were already ours; the two canvas spots are now sixty lines
each with no dependency. DECISIONS.md records the substitution.
The app installs, keeps a scan taken with no network in IndexedDB and
sends it when there is one, falls back to a page that explains itself,
and can push a deadline notice. Reading the log of what is queued is the
source of truth, so the notice clears when the capture actually lands.
The CSP now allows scripts by per-request nonce rather than by
'unsafe-inline'. That forced /offline to render per request: a
prerendered page carries a build-time nonce no live policy matches, so
its scripts were blocked and it never hydrated.
Two crashes fixed on the way. web-push throws on a VAPID subject that is
not https: or mailto:, and the code handed it APP_PUBLIC_URL, so any
machine with push keys died at boot; a misconfigured optional channel now
switches itself off and says why. And a subscription the push service
answers 410 for is deleted rather than retried forever.
Lighthouse on the production build: accessibility 100, best practices 96,
SEO 100, performance 73. The performance number is not trustworthy on
this machine and DECISIONS.md says why; total blocking time did fall from
17.6s to 1.7s once the hero canvas stopped drawing at full resolution
every frame and the landing page stopped importing GSAP.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The whole pipeline: storage behind one driver interface (local disk and S3),
a portable job queue with a poller, QR and CDC parsing, OCR through the
Anthropic API, dedupe, manual entry, and the bandeja that turns all of it into
one decision per card.
Scanning tries the trustworthy door first: a QR is parsed and prefilled from
its CDC; a photo without one is queued for OCR when a key is configured and
otherwise opens the manual form against the stored file. A QR that will not
parse records an ingest error and falls through rather than losing the photo.
Job claiming is the only dialect divergence, as SPEC allows: FOR UPDATE SKIP
LOCKED on Postgres, a conditional UPDATE against SQLite's single writer. Retry
backoff follows SPEC exactly and a job abandoned by a killed process returns to
the queue once its lock goes stale, which is the phase 3 acceptance case.
OCR uses structured outputs rather than parsing prose, so the model cannot
return anything but the RULES.md schema, and every field is nullable because
unreadable is a real answer.
Web: scan with live QR decoding (BarcodeDetector, ZXing fallback, wasm served
from our own origin), manual entry with the IVA split worked out from the
total, the bandeja with swipe, buttons and keyboard all doing the same thing,
and the documents list and detail with an editable classification.
The seed now carries Maria's 34 purchases and 8 sales and Carlos's 6, all
classified through the real rules, plus the two open ingest errors.
Two defects found and fixed with tests: seeded documents could be dated in the
future, which would corrupt any projection computed from them, and the category
buttons announced their keyboard shortcut as part of their name.
255 vitest tests, 37 Playwright tests, rules coverage still 100%, typecheck and
lint clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
