fix(knowledge): lint worker orphan self-heal did nothing on real data

[mkreyman]

What real-data verification found

Per the 'do the scheduled jobs actually work?' check, I ran KnowledgeLintWorker against the prod 77k corpus. Its orphan self-heal was effectively a no-op:

• 1,247 of 1,433 orphans have embeddings, but their nearest neighbor sits at ~0.55–0.59 — a near-miss of the global 0.6 link threshold. Re-linking at 0.6 produced zero links on every sampled orphan; they'd stay isolated forever.
• The other 186 orphans have no embedding, so ArticleLinkingWorker silently no-ops on them entirely.

Mocked unit tests passed; only real data exposed this — same lesson as the classifier bug.

Fix

• ArticleLinkingWorker accepts an optional per-job threshold (query-shaped arg, falls back to global config).
• KnowledgeLintWorker now splits orphans:
  • embedded → re-link at a lenient :knowledge_lint_orphan_link_threshold (default 0.5) so an isolated article connects to its closest relative instead of near-missing forever;
  • un-embedded → ArticleEmbeddingWorker (embeds, then chains to linking).
  • audit event now reports orphans_relinked and orphans_embedding_enqueued.

Tests

A near-miss neighbor (cos ≈ 0.55) links only with the lenient override; the lint worker embeds an un-embedded orphan and re-links embedded ones. Full local gate green (2999 tests, credo, dialyzer 0).

Note

0.5 is a contained default (orphans only — never global linking). Tune :knowledge_lint_orphan_link_threshold if you want stricter/looser orphan rescue.