Feedback weight closed-loop:
- search_hybrid_from_scored() accepts feedback_weight parameter
- retrieve_inner() loads per-user params from mem_user_retrieval_params
- Fulltext fallback path also uses per-user feedback_weight
- Hardcoded 0.1 replaced at all 3 scoring call sites
- .max().min() → .clamp() (clippy)

Observability:
- GET /metrics — Prometheus text exposition format
  memoria_memories_total, memoria_users_total, memoria_feedback_total,
  memoria_graph_nodes_total, memoria_graph_edges_total,
  memoria_snapshots_total, memoria_branches_total,
  memoria_async_tasks, memoria_governance_last_run_timestamp, memoria_info

Config visibility:
- GET /admin/config — runtime config with redacted DB password

Tests:
- test_tuning_affects_scoring rewritten to verify scoring math directly
  (avoids MatrixOne fulltext index flakiness on consecutive retrieves)
- Tool count tests updated: 15 → 18

Two processes with identical MEMORIA_INSTANCE_ID would previously
share the same holder_id, causing the re-entrant lock path to let
both processes acquire the governance lock simultaneously.

Fix: Config::from_env() appends the OS process ID to instance_id:
  MEMORIA_INSTANCE_ID=pod-0 → holder_id = 'pod-0-12345'

This ensures each process has a unique holder_id regardless of
the configured base name, while preserving re-entrant behavior
within a single process.

Tests:
- test_governance_daily_tunes_params_in_db: verifies DefaultGovernanceStrategy
  Daily task actually writes updated feedback_weight to mem_user_retrieval_params
- test_duplicate_instance_id_lock_is_exclusive: verifies PID suffix prevents
  two processes with same base instance_id from both acquiring the lock

tune_params() has three branches:
- useful_ratio > 0.7  → weight * 1.1 (was tested)
- negative_ratio > 0.5 → weight * 0.9 (NEW)
- neutral zone         → weight unchanged (NEW)
- weight clamped [0.05, 0.2] (NEW)

Added test_tune_params_negative_feedback_decreases_weight covering
all four cases via in-memory MockStore (no DB required).

Logging:
- Add TraceLayer: every request logs method+path+status+latency_ms
  (error level for 5xx, warn for 4xx, debug for 2xx/3xx)
- Auth failures now logged with warn! + token prefix (first 8 chars)

Input validation (DoS prevention):
- content: reject empty or >32 KiB
- top_k: clamp to [1, 100] on retrieve and search
- batch_store: reject >100 items; per-item content size check

Prometheus metrics (new counters):
- memoria_auth_failures_total: incremented on every 401/403
- memoria_sensitivity_blocks_total: incremented when sensitivity
  filter blocks a store request

Merged plan-memory-integration into goal-driven-evolution.md:
- Query memory before starting multi-step tasks (GOAL, LESSON, ANTIPATTERN)
- Register goals for multi-session work
- Track steps during execution (working type)
- Capture user feedback immediately (procedural type)
- Iteration review and cleanup workflow
- Changed inclusion from agent_requested to always

Applies to all AI tools: Kiro plan panel, Cursor Composer, Claude multi-step.

Merged plan-memory-integration into goal-driven-evolution.md:
- Query memory before starting multi-step tasks (GOAL, LESSON, ANTIPATTERN)
- Changed inclusion from agent_requested to always

Preserved all original content:
- 📋 PLAN structured storage with steps and risks
- 👍 FEEDBACK for positive experiences
- 🔄 RETRO with Completed [M/N], Key insight, Next improvements
- Bootstrap queries RETRO for context restoration
- Branch merge complete workflow (diff, checkout, merge, delete)
- Goal completion records Iterations count and Final approach

Added:
- When Goal is Abandoned section
- Don't create goals for quick fixes rule

Applies to all AI tools: Kiro plan panel, Cursor Composer, Claude multi-step.

- README: added memory_feedback to Core Tools table
- API Reference skill: added Feedback & Adaptive Retrieval section
  - POST /v1/memories/{id}/feedback
  - GET /v1/feedback/stats
  - GET /v1/feedback/by-tier
  - POST /v1/retrieval-params/tune
  - GET /v1/retrieval-params
- Steering rules (kiro/cursor/claude): added memory_feedback to Read tools
  with usage guidance (when to call, signal types)