feat(salience): Phase 1 — first-class standing tier (default-off, soak-gated)

[cipher813]

Summary

Ships the salience-tier substrate as a schema-backed, MCP-tooled, CLI-driven first-class feature. Memories explicitly promoted via memory_promote are injected into every <mnemon-context> envelope on every prompt, regardless of query similarity. Cap=15 (hard ceiling 20). Default-off behind STANDING_TIER_ENABLED.

Closes the salience-tier Phase 1 work per the 2026-05-22 reframing: Phase 1 IS the validation gate. Operator promotes ~5 career-context memories via the new memory_promote MCP tool, flips the flag, observes ≥1 week soak for runway-style under-weighting recurrence vs absence. The earlier synthetic-A/B-against-Phase-0 plan was reframed because the injection mechanism is identical between Phase 0's env-var-flagged form and Phase 1's schema-backed form — an A/B carried no marginal information.

Driver

The original Claude Desktop failure mode: load-bearing facts (Brian's runway, career posture) were retrieved into context but under-weighted against a dominant generic prior. RAG worked. Salience didn't. This PR closes the recall-side weighting gap. The upstream curation gap (capture fragmenting load-bearing facts into pieces) was closed by capture attention Phase A (#153, merged earlier today).

What ships

Store API

• promote_to_standing(id) — raises StandingTierCapReached / StandingTierProvenanceRejected / StandingTierError with user-actionable messages. Idempotent re-promote returns True.
• demote_to_situational(id) — round-trip. Idempotent on already-situational (returns False).
• list_standing() — live tier members, content-included, ordered DESC by created_at.
• standing_tier_status() — {count, cap, hard_ceiling}.
• search_bm25 + search_vector — gain include_standing: bool = False kw param. Tier 1 docs excluded from ranked retrieval by default (no double-counting; they're already injected unconditionally).

MCP tools (both stdio + Streamable HTTP)

• memory_promote(id) — surfaces cap / provenance / missing rejections as readable messages.
• memory_demote(id) — idempotent.
• memory_list_standing() — JSON array consumed by build_context.
• Tool count: 14 → 17.

CLI

• mnemon standing list / promote <id> / demote <id>
• mnemon status gains a Standing tier: N/CAP line.

build_context wiring

• When STANDING_TIER_ENABLED (config constant OR MNEMON_STANDING_TIER_ENABLED env override accepting 1/true/yes/on): single memory_list_standing round-trip → renders as ## Standing context sub-section inside the existing Layer 1 envelope, ahead of ## Situational recall.
• Phase 0 env-var path PRESERVED as fallback (MNEMON_STANDING_TIER_FILE → standing.json → standing-rendered.md cache). Operators retain a per-session override mechanism.

Schema

• documents.tier TEXT NOT NULL DEFAULT 'situational' — additive migration in _migrate_tier(). Index idx_documents_tier scoped to live rows. Pre-existing memories default to 'situational'; harmless if flag stays off.

Composability (all preserved)

• Layer 0 (is_well_shaped) — capture rejection runs upstream of any tier consideration
• Layer 1 envelope — standing block sits inside the same <mnemon-context> data-marking + nonce
• Layer 4 (HOOK_SOURCE_CONFIDENCE_CEILING + provenance) — hook-sourced memories cannot be promoted; explicit StandingTierProvenanceRejected rejection
• rc16 source_key upsert — unchanged; tier orthogonal
• Capture attention Phase A — recurrence_count signals candidates that operator-review can promote (bridges to salience-tier Phase 2 promotion signals work)

Soak gates (before flipping default-on)

1. ≥1 week with STANDING_TIER_ENABLED=true (set via env or config)
2. Observed reduction in runway-style under-weighting recurrence on real career-strategy conversations
3. Zero spurious-injection complaints from operator review of every promoted memory

Test plan

• All 22 new tests pass (pytest tests/test_standing_tier.py)
• Full suite 814 → 836 passing, no regression beyond the tool-count assertions in test_server_remote.py (bumped 14 → 17)
• CLI smoke: mnemon status shows tier line; mnemon standing list/promote/demote round-trip works on a temp vault
• Post-merge operator workflow: promote 5 career-context memories (ids 2543, 2084, 131, 2402, 2401 are the candidates surfaced 2026-05-22) via memory_promote MCP tool; set MNEMON_STANDING_TIER_ENABLED=true in the Claude Code launching shell; verify the <mnemon-context> block contains ## Standing context sub-section; observe ≥1 week soak.

🤖 Generated with Claude Code