feat(bot): episodic memory context in thinking partner

[Mathews-Tom]

Summary

Connect episodic memory to the thinking partner so past decisions surface alongside vault notes and graph context when thinking about a topic. Previously, the episodic store existed in isolation — episodes could be created and queried via bot commands, but the thinking partner never accessed them. Now, when a user thinks about "kubernetes scaling", relevant past decisions about kubernetes or scaling are automatically injected as context.

Inspired by memora-lab/memory-service-public's retrieval agent which pulls observations, people, and context into a unified prompt.

How It Works

User: /think kubernetes scaling strategy
                    ↓
ThinkingPartner._build_vault_context():
  1. Semantic search → vault notes about kubernetes
  2. Graph lookup → entity connections for kubernetes
  3. [NEW] Episodic search → past decisions mentioning "kubernetes" or "scaling"
                    ↓
Context sent to LLM:
  ## Relevant Notes from Vault
  ...
  ## Knowledge Graph Context
  ...
  ## Past Decisions                              ← NEW
  - **Adopt Kubernetes for deployment** → Reduced downtime by 50% (success)
    Lessons: Container orchestration simplifies scaling

Episodic Context Assembly

1. Extract topic words (>= 3 characters, first 5 words)
2. For each word, query EpisodeStore.search_by_entity(word, limit=3)
3. Deduplicate episodes across overlapping entity searches
4. Format as "Past Decisions" section with decision, outcome, status, and up to 3 lessons

Short words (< 3 chars like "ai", "ml") are skipped to avoid noisy broad matches.

Changes

Modified Files

• src/vaultmind/bot/thinking.py — Added episode_store: object | None = None parameter to think() and _build_vault_context(). New _add_episodic_context() helper queries episodes by topic keywords, deduplicates, and formats the "Past Decisions" section. Uses lazy EpisodeStore import with isinstance check to avoid circular deps
• src/vaultmind/bot/handlers/context.py — Added episode_store: object | None = field(default=None) to HandlerContext dataclass
• src/vaultmind/bot/handlers/think.py — Passes ctx.episode_store to ctx.thinking.think()

New Files

• tests/test_episodic_thinking.py (321 lines) — 9 integration tests using real EpisodeStore with tmp_path

Backward Compatibility

• episode_store defaults to None in all signatures — existing callers unaffected
• HandlerContext.episode_store defaults to None — existing handler construction unchanged
• When episode_store is None, no episodic section is generated (graceful skip)
• All existing thinking tests pass unchanged (no episode store configured)

Test plan

• 9 new tests in test_episodic_thinking.py:
  • Entity match surfaces episodes with decision, outcome, and lessons
  • No matching episodes → no "Past Decisions" section
  • Deduplication: same episode matched by multiple keywords shown only once
  • None episode store → graceful skip
  • Multiple episodes all shown
  • Pending episodes display "pending" status
  • Lessons truncated at 3 per episode
  • think() passes episode_store through to context builder
  • Short words (< 3 chars) skipped in entity search
• All existing thinking/session tests pass unchanged (8 in test_thinking.py)
• Full suite: 966/966 tests pass, 0 regressions
• ruff check — clean
• mypy --ignore-missing-imports — clean
• Manual: create episodes via /decide + /outcome, then /think about related topic — verify episodes surface