Discovery final review fixes: lean publish, robust scoring, cheap prune (BL-0051)

- Cache prune counts directory entries first and only reads files when over
  the cap; tests for cross-kind prune, schema-mismatch and corrupt entries.
- Scoring tolerates malformed manga.json and cached records (non-list tags,
  object or null relations, non-object items); a tag without a rank no longer
  passes the top-tag filter; authors without an integer AniList id are dropped.
- discovery.json publishes only candidates some feed shows; examined/clipped
  still describe the pool. Tests that asserted the published count equals the
  pool size now check the pool via candidatePool() and the published set
  against the feeds.
- The post-harvest rescore failure test now breaks discovery.json right before
  the publish (via a scraper stand-in) instead of relying on tags: 5, and a
  following successful harvest clears the stale lastError.
- Consent text names everything sent to AniList; paused label mentions admin
  harvests; the UI test asserts the XSS seed instead of skipping it.
- Slot-mix test proves the fresh/outside reservations matter.
- Docs: privacy note, fresh fallback reason, per-period cap, dry-run wording,
  published-candidate rule; plan/spec hygiene and spec status.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

docs/architecture.md

@@ -251,6 +251,8 @@ The reader alone talks to AniList for discovery ([discovery.md](discovery.md)).
   startedAt, finishedAt, requests, budget, stopReason }`. `stopReason` is `budget`, `rate-limited`, `forbidden`,
   `challenge`, `upstream-errors` or `error` (a local fault). A run cut off by a restart is marked `interrupted`
   at startup; queued refreshes are not persisted.
+- **Published candidates:** only those at least one feed shows. `lastRun.candidateCount` (the admin
+  candidate count) counts that set; `lastRun.examined` and `clipped` still describe the whole pool.
 - **Published fields:** candidates found through AniList carry `tagRecords`, `authorRecords`, `seed`
   (`recs|creator|tagTop|fresh`) and `fetchedAt`; feed items carry `group` (seed kind or period) and `reasons`
   (all reasons, strongest first) next to `reason`; seed evidence carries `fetchedAt`.