feat: verbatim transcripts — sessions + messages (refs #107)

[pszymkowiak]

Summary

Add a third memory system — Transcripts — alongside the existing curated memories and knowledge-graph memoirs. A transcript is a verbatim recording of an agent conversation: every user turn, every assistant reply, every tool call stored as-is, no summarization.

Search happens at query time via FTS5 BM25 (boolean, phrase, prefix), not at write time. This is the pattern validated by MemPalace's 96.6% LongMemEval score: raw > extraction for retrieval fidelity.

Closes the gap in issue #107 and positions ICM for three use cases it couldn't previously serve:

• Session replay — post-mortem review, debugging what an agent actually did
• Compliance / audit — SOC2, HIPAA, full trail (no lossy extraction)
• Training data — collect gold-labeled sessions for fine-tuning / RAG

Why this belongs in ICM (not a separate tool)

• Same SQLite file as memories/memoirs → one store, consistent backup
• FTS5 is already in the binary; transcripts reuse the same index engine
• Rust + SQLite write throughput is ~10× faster than ChromaDB-based alternatives and produces smaller on-disk files
• Zero new runtime dependencies — no Python, no ChromaDB, no external service

Scope of this PR

• Schema — `sessions` + `messages` tables, FTS5 virtual table with triggers, ON DELETE CASCADE
• Core types — `Session`, `Message`, `Role`, `TranscriptHit`, `TranscriptStats`, `TranscriptStore` trait
• Store — 7 `SqliteStore` methods (create, get, list, record, list messages, search, forget, stats)
• CLI — `icm transcript {start-session, record, search, list-sessions, show, stats, forget}`
• MCP — 5 new tools (total MCP surface: 22 → 27)
• Tests — 8 unit tests covering FTS5 boolean/phrase, scoping, stats, cascade delete, ordering
• README — new section under `## CLI` with end-to-end example

Not in this PR — explicit follow-ups

• `fix: update README license from MIT to source-available #2` TUI tab "Sessions" in `icm dashboard`
• `fix: improve extract fact scoring for dev tool outputs #3` Web dashboard route `/sessions` + `/sessions/:id` thread viewer + REST endpoints
• `Add tests and documentation #4` `icm hook transcript` — auto-capture from Claude Code / Gemini / Codex hooks
• `Add performance tests, improve docs #5` Optional RTK Cloud sync for team-tier paid audit / replay

Splitting keeps each PR reviewable (this one is already ~1.5k lines).

Manual smoke test (reproducible)

```bash
SID=$(icm --db /tmp/test.db transcript start-session -a claude-code -p demo)
icm --db /tmp/test.db transcript record -s $SID -r user -c "Postgres ou MySQL ?"
icm --db /tmp/test.db transcript record -s $SID -r assistant -c "Postgres: JSONB natif, BRIN indexes."
icm --db /tmp/test.db transcript record -s $SID -r tool -c '{}' -t Bash --tokens 42
icm --db /tmp/test.db transcript search 'postgres OR mysql'
icm --db /tmp/test.db transcript search '"BRIN indexes"'
icm --db /tmp/test.db transcript show $SID
icm --db /tmp/test.db transcript stats
```

Test plan

• `cargo test -p icm-store transcript` → 8/8 pass
• `cargo build --release -p icm-cli` → clean
• Manual end-to-end smoke via CLI (see above)
• Reviewer: try FTS5 syntax edge cases (prefix `postg*`, negation `postgres NOT mysql`, proximity `NEAR(brin json, 5)`)
• Reviewer: confirm `ON DELETE CASCADE` behaves correctly across SQLite versions

[pszymkowiak]

wshm · Automated triage by AI

📊 Automated PR Analysis

✨ Type	feature
🟡 Risk	medium

Summary

Adds a new 'Transcripts' subsystem to ICM for verbatim session replay, including new SQLite tables (sessions + messages with FTS5), a TranscriptStore trait with 7 methods on SqliteStore, 7 CLI subcommands, 5 new MCP tools, and 8 unit tests. This enables storing raw agent conversations for replay, audit, and training data use cases.

Review Checklist

• Tests present
• Breaking change
• Docs updated

Linked issues: #107

---

Analyzed automatically by wshm · This is an automated analysis, not a human review.