rdmsm4x - changelog - RDReceipt receipt inventory 1641 unique files labeled and tagged - claude - TASK-20260926-73 - RDReceipt - 20260926-2311

Inventoried and labeled every receipt candidate with RDReceipt's pipeline; fixed a parser date bug; Finder-tagged 4,835 source files in place (no content/mtime change). Notes entry PENDING: headless Background session.

RESULT — TASK-20260926-73: RDReceipt receipt inventory

[2026-09-26 23:10:21 EDT · rdmsm4x] · run 22:57:51 → 23:10 EDT · claude@rdmsm4x (headless, session expenses-receipt-inventory-20260926) · worktree ~/dev/_worktrees/rdreceipt-inventory-20260926, branch expenses/rdreceipt-inventory-20260926

The answer Rich wanted: current expense state

• Inside Arista's 90-day window (since 2026-06-28), the only unsubmitted business items are recurring bills. The claimable ones are 3 Starlink invoices and 3 AT&T bills: $860.00 (AT&T capped at $110 per the draft's convention). That is proposed report R001. There are also 2 Verizon Wireless Business payments ($881.60; a third on 9/01 is a near-duplicate of the 8/27 one). Rich's draft marked those "borderline, claim $0", so they are listed but not grouped.
• No one-off receipt or invoice (meal, parking, travel) exists in 2026-06 through 2026-09. Either there were none, or they never reached these folders. The newest one-off receipts are Eataly on 6/15 (reservation only) and Joe's Ginger on 5/18.
• Everything older than the window, newest first: R002 (2025-12-09 → 2026-06-25, $2,942.62) and R003 (2025-09-10 → 2025-12-09, $2,937.66) cover the same ground as the draft's Batch 1 and Batch 2. Since 2024-08-17 (after the last Concur package), 84 grouped items come to $10,894.28. All of them are past 90 days.
• In all, 21 proposed reports (each ≤ $3,000, or a single item over the cap, flagged) hold 245 items worth $62,926.01. Most of the pre-2024 items are old enough that submission is doubtful. They are grouped so they're visible, and each group is flagged "older than 90 days".

Outputs (all in ~/dev/finance/expenses/inventory/)

Finder tags kind, YYYY-MM, category and submitted|unsubmitted, plus xattr com.eastcoastscience.rdreceipt.label (JSON), are on 4,835 on-disk files, with 0 errors. The 594 zip members are not tagged, and every one of them also exists as a loose file.

Counts

Positive control

~/Documents/_receipts/_q3-2024_receipts/24_09_starlink_092524_INV-USA-29503623-76769-94.pdf → kind=recurring bill, date=2024-09-25, merchant=Starlink, total=120.00 USD, tax=0.00, itemized=yes, category=phone/internet, confidence=0.85, unsubmitted, within_90_days=no, group R005. Its Finder tags as read through URLResourceValues.tagNames (the API Finder uses): ["recurring bill", "2024-09", "phone/internet", "unsubmitted"]. All fields match the invoice text (Invoice Date September 25, 2024; Total Charges USD 120.00; Total Tax USD 0.00).

This control caught a real bug. Before the fix it read 2026-09-25. The invoice's year wraps to the next line, and NSDataDetector fills in the current year, which would have put a two-year-old bill inside the 90-day window. The fix is in the shipped parser, commit 9156fe3.

Second control, in the window: bills/starlink/starlink_3764_2026-09.pdf → 2026-09-25, Starlink, $140.00, R001, matching the draft's Batch 1 line exactly. Negative control: rich_arista_W2_2017_TAX.pdf → not-a-receipt with no total. Before the tax-form rule it was a "receipt" dated today for $116,250.

Verification that nothing was modified

• After tagging, all 4,835 on-disk source files were re-hashed and re-stat'ed against the pre-tag backup: mtime changed 0, size changed 0, sha256 changed 0.
• The 5 ~/Documents/arista_expenses*.zip files were read only (extracted to /tmp/rdr-inv-20260926/zips), and their sha256 values are unchanged (6f915e02…, cc2073c6… ×2, 8b809263…, b92827ff…).
• ~/dev/arista and ~/dev/apps/autoCE were never opened. The harness refuses them (a ~/dev/arista root exits 2 with REFUSED), and the labeler asserts that no manifest path contains either one.
• Re-running the labeler is idempotent: on the second run all 4,835 tags came back unchanged.
• Undo: python3 ~/dev/_handoff/rdreceipt-inventory-20260926/untag.py restores each file's pre-run tag xattr (none existed) and removes the label xattr. Tested on a scratch copy, where tags went from 4 to none. A --dry-run reports 4,835.

Code changes (worktree, committed by explicit path, not merged, not pushed)

"ReceiptScannerKit" does not exist in this repo. The pipeline is Apps/ReceiptRoo/Shared/Engine, and that is what ran.

Decisions made (DEC-style, reversible)

1. Submitted = a byte copy in _archive_submitted060424/, Receipts0816/, or Dropbox Receipts (ALL)/2024-08/. The Dropbox folder is a copy of the Aug 16, 2024 Concur package: same filenames, 72 byte-identical files, created 2024-08-21. A near-duplicate of a submitted file inherits submitted. Measured effect: the Dropbox folder alone changed no row, because each of its unique files also has a copy in one of the brief's two folders. Near-duplicate propagation raised submitted from 110 to 181. The Dropbox rule stays in as a guard for future files.
2. likely-submitted is a CSV status only (dated ≤ 2024-08-16 and filed in an expense-archive folder, or carrying the 06/08/2023 export filename). The Finder tag stays unsubmitted for these. Only proven submissions get the submitted tag, so a missed claim can't hide behind a guess.
3. Store location: a dedicated rdreceipt-inventory.sqlite3 using the app's schema, not the app's live Application Support store. The SQLCipher vault swap (STATUS item 6, DEC-20260901-02) is still pending, and writing thousands of rows into the plain store now would create data that has to be migrated.
4. Report grouping follows Rich's draft conventions: AT&T is claimed at a $110 cap, and Verizon and Google Fi are kept out of groups ("borderline, claim $0"). Groups split at the 90-day line so a within-window report never mixes with stale items.
5. Sanity guards: a date later than the file's own save time is rejected. A total over $10k is dropped and noted in sanity_notes. Scan-timestamp filenames are used only as a 0.3-confidence fallback, flagged "date is the scan date".

Known limits

• These labels are heuristics. The 90-day-window rows were checked by hand against the draft. Older rows were not. The 107 rows under 0.5 confidence are listed in summary.md.
• "Itemized" means the parser found at least 2 line items. It undercounts on photographed receipts, so the 287 "meals missing an itemized receipt" flags are an upper bound.
• .txt/.eml bill transcriptions in bills/ are inventoried as not-a-receipt (a file-type rule) even though they record real bills.
• The Apple Notes changelog is pending: this headless session is launchctl managername = Background, and notes_changelog.zsh refuses to run from that. The file copy is in ~/dev/LLM/Claude/changelogs/.

Verbatim commands

S=/tmp/rdr-inv-20260926
unzip -q -o ~/Documents/_receipts/arista_expenses.zip -x '__MACOSX/*' -d $S/zips/zip-cc2073c6
unzip -q -o ~/Documents/arista_expenses_folder.zip   -x '__MACOSX/*' -d $S/zips/zip-8b809263
cd Apps/ReceiptRoo
swiftc Tools/inventory/InventoryHarness.swift Shared/Engine/ReceiptScanner.swift Shared/Engine/FileHasher.swift \
  Shared/Engine/ReceiptReader.swift Shared/Engine/ReceiptFieldParser.swift Shared/Engine/ReceiptClassifier.swift \
  Shared/Engine/ReceiptDatabase.swift Shared/Models/Receipt.swift -o /tmp/rdreceiptInventory -parse-as-library -lsqlite3
/tmp/rdreceiptInventory ~/dev/finance/expenses/inventory/rdreceipt-inventory.sqlite3 $S/manifest.json \
  ~/Documents/_receipts "$HOME/Library/CloudStorage/Dropbox-richh/Receipts (ALL)" $S/zips/zip-cc2073c6 $S/zips/zip-8b809263 \
  ~/dev/finance/expenses/receipts ~/dev/finance/expenses/bills ~/dev/finance/expenses/source_files
#  → DONE files=6171 ingested=1183 resumed=0 dupes_in_run=4464 failures=0 excluded_entries=0 store_rows=1183  (23:00:31→23:04:36)
/tmp/rdreceiptInventory2 --reparse ~/dev/finance/expenses/inventory/rdreceipt-inventory.sqlite3   # built from 9156fe3
#  → REPARSED rows=1183 dates_changed=29
/tmp/receiptFixtureCheck ../../core/fixtures/rdRECEIPT_field_parsing.json ../../core/fixtures/rdRECEIPT_document_classification.json
#  → fixtures: 17 cases across 2 file(s), all pass
cd ../..
uv run tools/inventory/label_inventory.py --store ~/dev/finance/expenses/inventory/rdreceipt-inventory.sqlite3 \
  --manifest $S/manifest.json --outdir ~/dev/finance/expenses/inventory \
  --zipmap "$S/zips/zip-cc2073c6=<5 identical zips, | separated>" --zipmap "$S/zips/zip-8b809263=$HOME/Documents/arista_expenses_folder.zip" \
  --drafts ~/dev/finance/expenses/expenses.csv --today 2026-09-26 --tag
#  → unique=1641 groups=21 eligible=245 tags {changed/unchanged: 4835, zip-member-skipped: 4008} tag_errors=0 hash_errors=0

Evidence files in this folder: ingest-run.log, manifest.json (paths, sizes and hashes only, with no OCR text), pretag-xattr-backup.json, and untag.py.