fix(extraction): read stickers at any orientation, and flag what it can't

[iiamit]

Customer report: pages carrying an upright sticker and one rotated 90° came back with only the upright entry.

The rotated entry wasn't wrong — it was absent. And an absent entry has no low-confidence field to catch the eye, so the owner had no way to know it was missing. That's a worse failure than a misread value, and it's the part this PR is really about.

Why it happened

Three things in EXTRACTION_SYSTEM_PROMPT combined:

1. Stickers were described only as content to merge into a single entry — never as entries in their own right.
2. "return one object per entry, in top-to-bottom order" bakes in one vertical reading order. A sideways sticker has no slot in it.
3. Orientation was never mentioned at all.

Stage 1 — prompt (no extra cost)

States that shops affix stickers wherever there's room, at any orientation (90°, upside down, angled); that each sticker is normally its own entry; and that ordering never justifies omitting one.

Stage 2 — make a miss visible

The model now reports unread_rotated_content when it can see rotated content it couldn't fully read. It's required in the schema, so it can't quietly default to false.

That triggers one targeted second pass, and if that still comes up short the page keeps the flag and Review shows an amber warning pointing at "Add an entry the extractor missed".

The second pass re-reads the same bytes with a narrower prompt rather than rotating the image. A vision model reads rotated text when told to look — and re-encoding would mean importing sharp into app code, which package.json deliberately avoids: it's pinned for a CVE precisely because user-uploaded images reach it via image optimization, and the comment there notes that nothing in the app imports it.

Gated on the model asking, so the extra call lands on the few pages that need it rather than tripling the cost of every page against the per-user daily cap and global dollar ceiling. Best-effort throughout — a failed retry never costs the first pass's entries.

The bit most worth reviewing

stillUnreadAfterRetry() is a separate pure function with its own test, because inverting it would suppress the warning on exactly the pages that need it:

return second.unread_rotated_content || second.entries.length === 0;

An empty retry is not evidence the page is clean. It errs toward keeping the warning.

Migration

0052_page_rotated_content.sql adds page.unread_rotated_content, default false — documented as "nothing to flag", not "verified clean", since pages extracted before this were never assessed.

Apply to prod and test before merging. 462 unit tests, typecheck, lint and build clean. /help updated.

Worth asking the customer

A photo of one of those pages would be worth more than my reasoning about which stickers get missed — and it becomes the regression fixture. Also worth telling them "Add an entry the extractor missed" already exists on the review screen as a manual escape hatch today.

https://claude.ai/code/session_01XBNGwWrPih2Xgu6MVrcd6R