Add Phase 3 retrospective (miroir-r3j)

Documents lessons learned from implementing the 14-table task store:
- What worked: migration-first approach, trait abstraction, property tests
- What didn't: initial schema design, manual pruning
- Surprises: rusqlite JSON handling, Redis async/sync bridging
- Reusable patterns for multi-backend store implementations

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

notes/miroir-r3j-retrospective.md

@@ -0,0 +1,45 @@
+# Phase 3 Retrospective — miroir-r3j
+
+**Task:** Task Registry + Persistence (SQLite schema, Redis mirror)
+**Status:** Complete
+**Date:** 2026-05-05
+
+## Summary
+
+Implemented the full task store with SQLite and Redis backends covering all 14 tables from plan §4.
+
+## What Worked
+
+1. **Migration-first approach** — Creating migrations first with the schema_migrations module made version tracking straightforward and idempotent.
+2. **Trait abstraction** — Defining TaskStore trait once and implementing it for both SQLite and Redis kept the APIs consistent.
+3. **Property tests with proptest** — Found edge cases in JSON serialization (node_tasks, node_errors, history arrays) that unit tests wouldn't have caught.
+4. **Redis index sets** — Using `*_index` sets for O(1) listing avoided SCAN and made list operations predictable.
+
+## What Didn't
+
+1. **Initial schema design** — Had to add `node_errors` field in migration 003; better to have included it in 001_initial.sql.
+2. **Redis EXPIRE vs manual pruning** — Initially tried manual pruning for sessions/idempotency, but Redis EXPIRE is simpler and more reliable.
+
+## Surprises
+
+1. **rusqlite JSON handling** — No native JSON type; had to serialize/deserialize manually for node_tasks and history fields.
+2. **Redis async in sync context** — Needed `block_on` helper with dedicated thread runtime to avoid "runtime inside runtime" panics.
+3. **SQLite subquery DELETE LIMIT** — Not directly supported; needed rowid IN (SELECT rowid ... LIMIT) pattern.
+
+## Reusable Pattern
+
+For future multi-backend store implementations:
+- Define trait first with all CRUD operations
+- Use migration registry pattern (versioned SQL files)
+- For Redis: use `table:_index` sets for listing, avoid SCAN
+- For SQLite: use WAL mode + busy_timeout for concurrent access
+- Property tests are worth the setup for JSON-heavy schemas
+
+## Deliverables
+
+- SqliteTaskStore (2551 lines) with all 14 tables
+- RedisTaskStore (3939 lines) with all 14 tables  
+- 3 migration files with version tracking
+- REDIS_MEMORY_ACCOUNTING.md (305 lines)
+- Helm values.schema.json enforcing replicas > 1 → redis
+- 12 Phase 3 DoD tests + 36 SQLite unit tests