rdmsm4x-changelog-20260831-2148-rdreceipt-ocr-deadlines-corpus-run-publication-and-continuity

rdmsm4x — RDReceipt: OCR deadlines, full-corpus run, publication, session continuity

2026-08-31 12:45 → 21:48 EDT on rdmsm4x. Claude Code, Opus 5. Continues the 12:44 changelog.

One line: RDReceipt gained a hard OCR timeout, was run over its whole 3,353-file corpus for the first time with zero failures, got its first page on dev.ecs0.net (removing a false verified claim that had been published for a week), and was registered and made resumable under the new SESSION_CONTINUITY standard.

Scope

One project: ~/dev/apps/RDReceipt, worked in its worktree .claude/worktrees/rdreceipt-chat-thread-setup-a94cd2. Fleet files touched: ~/dev/PROJECTS.md, ~/dev/data/products.json (+ status-updates/), ~/dev/data/project-docs/RDReceipt.json, ~/dev/fleet/ops/resume/*-rdreceipt*, and one new LaunchAgent. No other host touched.

What changed

File	Change
`Apps/ReceiptRoo/Shared/Engine/ReceiptReader.swift`	per-page (60 s) and per-file (180 s) OCR deadlines; `timedOutPageCount`; `-partial` engine suffix
`Apps/ReceiptRoo/Shared/Engine/IngestEngine.swift`	counts timed-out pages; engine `switch` made prefix-matched so a `-partial` read is not filed as "hybrid"
`Apps/ReceiptRoo/Tools/check_fixtures/FixtureCheck.swift`	now runs both fixture files, routed by version key
`CORPUS-RUN.md` (new)	the 3,353-file run, led by what the numbers are not
`STATUS.md`, `ISSUES.md`, `SESSION-STATE.md`, `CLAUDE.md`	items 7, 8, 11 closed; state rewritten to the six-question standard
`~/dev/data/project-docs/RDReceipt.json` (new)	source for the published page; 17 features with boundaries
`~/dev/fleet/ops/resume/{PINNED-SESSION-rdreceipt,resume_rdreceipt_on_login.zsh}` (new)	auto-resume for this session
`~/Library/LaunchAgents/com.eastcoastscience.claude-resume-rdreceipt.plist` (new)	loads it at GUI login

Evidence

Corpus run: 3,353 files in ~17 min, zero read failures, zero OCR deadline hits. 1,037 distinct after content-hash dedup — 2,316 files (69.1%) are exact duplicates. Of 868 distinct PDFs, 25.7% needed no OCR, which finally sources the "26% of PDFs" figure that had lived only in products.json with no stated sample.
OCR deadline proven both ways: forced at 1 ms → vision-ocr-partial, no fabricated fields; non-overshoot re-verified at load average 24 → zero field differences vs the pre-deadline run.
Fixtures: 16 cases across 2 files, all pass; both branches made to fail against doctored expectations first.
Build: xcodebuild -scheme ReceiptRoo-macOS → ** BUILD SUCCEEDED **.
Page live: curl -H 'Host: dev.ecs0.net' http://127.0.0.1:8788/projects/rdreceipt.html → 200, zero external references. Checked docker inspect ecs0-dev-wiki-1 first to confirm which docroot is actually mounted — gen_site.py writes a different tree.
Resume agent: three paths verified — normal, qualifier-disabled (recovers, logs GLOB QUALIFIER FAILED), stale pin (stops, says so differently).

Corrections made to previously published claims

products.json published "installed on both iPhones" as tested-with-evidence. It is false — the app has never been built for a device. Removed; both dashboards now state the negative.
The product entry was renamed ReceiptRoo → RDReceipt (slug rdreceipt).
Two of my own claims about a peer's script were retracted after measurement; see below.

Backups / how to undo

Nothing deleted. ~/dev/data/products.json backed up beside itself (.bak-20260831-*). The superseded dashboard page was moved, not deleted: ~/dataroo.net/wiki/p/receiptroo.html → ~/dataroo.net/wiki/archive/superseded-20260831/. All code is on work/expense-export-and-parser-bakeoff at 87de52e, pushed to the private remote. Disable auto-resume with launchctl unload -w ~/Library/LaunchAgents/com.eastcoastscience.claude-resume-rdreceipt.plist.

Traps worth carrying off this project

A timeout measures the machine, not the file. A 20 s OCR deadline truncated a healthy page at load average 24. Verify a bound under the load that will actually exist.
A guard that inherits the option it depends on is not a guard. Proven by running from a shell with the option switched off, not by reasoning.
A recovery path the failure prevents you from reaching is not a recovery path. zsh aborts on an unmatched qualified glob; the fallback never ran until the glob was contained in a subshell.
Logged where nobody reads it is not the same as caught.
A fill rate is not an accuracy rate. 42 of 42 confidently-classified non-receipts still got a merchant.
Retracted after measurement: an unrecognized modifier error I attributed to a peer's script line — launchd.err.log (15:47:22) predates the file itself (15:47:44), so it came from a version that no longer exists.

Outstanding owner actions

DEC-20260831-05 — may receipt contents leave the machine via the planned MCP server?
DEC-20260831-02 — off-host backup of the 1.5 GB corpus (code is already on a private remote).
Merge to main is still blocked — the primary checkout holds another session's uncommitted dataroo→ecs0 rename in the same two files, unchanged for 46h.
Concur field recon remains blocked at the sign-in wall.

No secrets are recorded here. Work registered as FEAT-20260831-02; resumable state is that repo's SESSION-STATE.md.