Phase 9: efficacy attribution + reversible pruning — M-C (closes the loop)

[macanderson]

Phase 9 — efficacy attribution + reversible pruning (M-C). Closes the loop.

context_use / context_use_feedback with use_trace_id; the derived context_health_projection; reversible pruning (mark stale → stop auto-selecting → notify → archive after grace; never touch protected/blocking/user-confirmed directives). Attributable, opportunity-aware, trusted-method efficacy — agent_self_report alone cannot drive pruning. Review age produces review_due, never stale without efficacy/contradiction evidence.

The system now demotes advisory context that repeatedly doesn't help. Full Observatory receipts + the six-label status table can come incrementally after the core counters are right.

Blocked by Phase 6 and Phase 1 installment 2. Part of Epic #469.

[macanderson]

Triage 2026-07-26 — a competing efficacy loop is already live and must be reconciled, not duplicated

None of the specified design exists. context_use/context_use_feedback are taxonomy strings only (stella-core/src/context_record/kind.rs:46-47); use_trace_id is a struct field in the inert context_use.rs:222,241, which defers identity resolution at :272 ("context_use_id is deferred to the repository"). There is no context_health_projection (zero hits). The six-label status table is absent even as a type: RecordStatus is {Active, Retracted, Archived} (kind.rs:219), EffectiveStatus is {Active, Superseded, Retracted, Archived, Expired} (:242) — neither carries stale or review_due.

But an older, differently-shaped efficacy loop already runs and achieves part of this phase's goal:

• memory_citations + fold_citation_stats → positive_streak, negatives, truthful_rate (stella-store/src/lib.rs:307-326, :929).
• Demotion exists: quarantined: negatives >= QUARANTINE_NEGATIVES_THRESHOLD excludes a memory from recall entirely; only stella memory unquarantine clears it (lib.rs:321-325) — reversible, which is this phase's spirit.
• Reversible tombstones exist: the forgotten table (stella-store/src/ddl.rs:241) behind stella memory forget/restore (feat(memory): stella memory forget/restore, as a tombstone that survives re-learning #671), widened to episodes in context: stop recency from filling recall with unrelated frames #677.
• A stale label exists but on a different axis — path-anchor rot, not efficacy (stella-cli/src/memory_cmd.rs:277-320).

The critical conflict: cite_memory is an agent self-report, and it alone already drives both promotion and quarantine — which is exactly what this issue says must not be allowed to drive pruning.

So Phase 9 cannot be built beside this loop; it must replace or subsume it, or the repo ends up with two efficacy authorities disagreeing about the same memory. Worth stating in the body before anyone starts.

[macanderson]

Superseded by the four-phase adaptive-context plan merged in #692 (docs/design/adaptive-context-plan.md), which replaces the eleven-phase roadmap and the 2026-07 spec bundle this issue was written against.

Disposition — old Phase 9 (efficacy attribution + reversible pruning): carried forward largely intact as Phase 4 (#715).

Two changes worth knowing: the shipped citation loop is generalized into one evidence source among several rather than rebuilt, and negative attribution now requires that the item had an opportunity to help with the assessment method recorded — absence of citation is not evidence of uselessness.

Closing as superseded rather than abandoned — see the old→new table in the plan's Tracker reconciliation section. Epic #469 has been rewritten around the four phases.