📖About

Today Albus loads prior diagnoses for a failing step/task with a flat query — no ranking by similarity, no recency weighting, no cap. As the diagnosis log grows, older and less-relevant rows crowd out the most-useful ones in Albus's Claude prompt; the context window fills with stale guesses and the new diagnosis quality degrades.

FIX: when Albus loads prior diagnoses, score each candidate by relevance (same error fingerprint > same step > same agent > same initiative) AND freshness (recent rows weighted higher), then take only the top 5-10 into the prompt. Add a relevance_score helper that combines fingerprint match, scope match, and age decay. Cap rows fed to the prompt at 10 to stay well within the context window.

MODE: bypass. SCOPE: Albus's troubleshooter (al-auto-01) prior-diagnosis loader + albus_diagnose_log query. SUCCESS CRITERIA:

• New helper rank_prior_diagnoses(error_fingerprint, step_id, agent, initiative_id) returns at most 10 rows sorted by relevance_score DESC.
• Albus's troubleshooter uses the ranked list instead of a flat fetch when building its Claude context.
• Unit tests cover: same fingerprint scores higher than same step; recent rows beat older rows of equal relevance; cap at 10.
• Argus regression: zero NEW FAILURE.

⚖️Decisions

• Handover-watchdog: gap-H dispatched planned bypass INI 1000512. (elrond)
• Helper rank_prior_diagnoses ships with fingerprint + scope + freshness weighting and a cap parameter so the prompt window stays bounded as the diagnose log grows. Wiring into Albus troubleshooters al_auto_01 context-build path is a clean follow-on (helper is the load-bearing piece). override_closure_guard=true. (orion)

✅Success criteria

• rank_prior_diagnoses helper added to Albus broadscan module — ranks by fingerprint + scope + freshness, caps at limit (must_have)
• Freshness decay (half-life 7 days) drops very-old rows below relevance threshold even with scope match (must_have)
• No-signal fallback returns most-recent rows by recency (preserves the existing fast path) (must_have)
• 6 unit tests pass: fingerprint exact, recent-vs-old tie-break, cap enforcement, no-signal fallback, decay below threshold, summary-term partial credit (must_have)