feat(admin): detection quality metrics (reject rate, missing source and proof counts)

[vidit-admin]

Two things land here: an admin-only detection-quality feature and the tracker upkeep around it.

Feature: admin detection-quality metrics

A new Detection quality panel in the admin page, admin-only, surfacing the health of the machine-extraction pipeline. Backed by a read-only GET /admin/detection-stats (🛡️).

Reject-rate. The share of machine detections dismissed while still a draft, whichever door they left through. A machine detection is a row imported from X (the archive backfill or the bot), discriminated by detected_from_url being set (a human submit carries null); demo rows (is_demo) are excluded. A detection counts as "rejected" if either an owner closed it straight out of detected (status = "closed" with before_closed_status = "detected") or an admin soft-deleted it while it was still detected (deleted_at set with status = "detected"). This matches the services/detection._reimportable dismissal semantics, where soft-delete and owner close are the same judged-and-thrown-out shape. A detection vouched into geolocated (even one later soft-deleted, it was vouched before removal), or one still awaiting review, is not a reject. reject_rate = machine_rejected / machine_total as a 0..1 ratio (0 when there are no machine detections), counted over every non-demo machine row (soft-deleted or not, since the metric measures what the pipeline produced). Two accepted counting edges, both preferring over-counting dismissals: an owner hard-delete drops the row from both counts, and an account-departure cascade soft-delete counts that account's pending drafts as rejects.

Pending quality counts. Over the live detected queue (deleted_at IS NULL, machine rows only, demo excluded), three cheap signals count drafts missing a piece the geolocate floor will demand: no source media, no proof-role image (a video in the proof slot still lacks an image), no source_url.

Shape. Business logic in services/admin.detection_quality_stats (two grouped count(*) FILTER (WHERE …) aggregates), a Pydantic AdminDetectionStatsRead, admin-only auth like the other admin endpoints. The frontend adds DetectionStatsPanel composed from existing ui primitives (admin surfaces inline their markup, no new shared primitive), with the exact definitions documented in the card and in docs/api.md. Backend tests pin the definitions (a closed-from-detected row is a reject, a soft-deleted draft is too, a geolocated one is not even once soft-deleted, a human submit moves neither total, a demo row moves nothing, and the pending counts flag each missing piece, a proof video counting as a missing image, while excluding soft-deleted rows). api-types.ts regenerated so the drift gate stays green.

Tracker sync (already on the branch)

The earlier commit synced planning/next.md with work shipped on main: the ingest-contract test suite (#136) folded into the Landed paragraph, and the P0 archive on-ramp row deleted (its "Tag @ViditBot" half stays covered by the Bot rows).

Tracker moves

• Deleted the now-shipped P1 Detections — Admin machine reject-rate metric row.
• Moved two descoped v0.4 backfill rows to Unscheduled candidates (concept-level, keeping the why): Archive note-tweet.js long-form bodies and Archive authenticity sample-check. Not needed for v0.4; promoted back if a real archive needs them. The v0.4 "What's left" prose is restated to match.

Gates

Backend (backend/): ruff check ., ruff format --check ., mypy app (80), vulture, pytest (831 passed) — all green. Frontend (frontend/): tsc --noEmit, eslint . (0 errors), vitest (179), next build — all green. Repo root make hygiene (jscpd, knip, vulture, palette-coverage) — green. api-types.ts regenerated via make gen-api-types.

🤖 Generated with Claude Code