meter-closeup: follow-ups from Codex review 5328852981 #103

These findings from Codex review 5328852981 on #16 were checked and are real. None changes a committed result or puts data at risk, so they wait until after #16 merges, per the manager's rule for this round.

• Sweep inputs in the cache key (sweep.py). Rows are validated against the label and the code digest, not the image bytes or the clean-read box. Replacing a photo at the same size would reuse old rows. Add an input digest.
• make q45 on a fresh data directory (locate.py). scan() opens ocr/scan.jsonl without creating ocr/; only make q1 creates it. Create the folder before writing.
• Joining OCR lines needs a horizontal gap limit (match.py). Any two observations on one row are joined, however far apart. On the committed data, 7 agreed reads use a joined row; each joins to exactly the labelled number, so none is a false read. Limit the gap to a few line heights anyway.
• Class labels by token, not substring (match.py). CL20 would match inside CL200. The one CL20 photo holds a genuine CL20 token, so the 20/22 result stands. Match a whole token.
• Stale review verdicts (review.py). Saved verdicts are keyed by photo ID only, so a changed label keeps an old "keep". Store the reviewed number with each verdict and drop mismatches.
• make q1 crash on a split read with no four-digit line (clean.py). tallest_digit_line can return None when the matched number is joined from short observations. The committed run does not hit it; guard it.

🤖 Generated with Claude Code