The analysis admin now reflects the OCR backend instead of the vision
LLM. Backend response types are unchanged (vision fields ride along,
ignored) so this is fully reversible.
- Settings → Analysis: keep only the OCR-relevant knobs (workers, job
timeout, max image bytes) plus a read-only "Backend: OCR (ocrs)"
line; drop the endpoint/model/API-key, sampling, slicing, and prompt
fieldsets.
- Analysis tab: reword the lede to OCR text extraction; the page-detail
modal drops the NSFW/content-warning, scene, and auto-tag sections and
shows only OCR text (the per-line kind label renders only when set,
since ocrs leaves it blank).
- History table: remove the NSFW-only filter and the per-row NSFW tag,
and stop sending the `nsfw` query param.
- Metrics panel: remove the by-model table (OCR has a single engine);
keep the tiles and trend charts.
Add component tests asserting the history table omits the NSFW filter
and never tags a row, and the metrics panel renders tiles without the
by-model table.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
89300605a4