feat: cross-session memory persistence with hybrid search (closes #133)

[luceinaltis]

Summary

Closes #133.

Wires up the 3-tier memory system so conversation context survives across REPL sessions.

Tier 2 — disk persistence (qracer/conversation/engine.py)

• ConversationEngine now takes an optional summaries_dir kwarg.
• When set, _maybe_compact() calls SessionCompactor.compact_and_save() instead of the in-memory compact(), writing ~/.qracer/summaries/<session_id>.md after the turn log crosses the 8 000-token threshold.
• After a successful save, if memory_searcher is also set, the summary is auto-indexed into Tier 3 via MemorySearcher.index_summary(session_id, summary) so the very next session can find it.
• Backward compatible: without summaries_dir, behaviour is unchanged (calls compact()).

Tier 3 — hybrid search (qracer/memory/memory_searcher.py)

• New embedding_fn: Callable[[str], list[float]] | None constructor parameter. Leave it None for keyword-only search (existing behaviour); pass any callable (Claude API, sentence-transformers, a stub, …) to get hybrid search.
• New session_embeddings (session_id VARCHAR PK, embedding FLOAT[], indexed_at TIMESTAMP) table stores vectors alongside the existing FTS session_index.
• _vector_search() runs cosine similarity via DuckDB's native list_cosine_similarity, joining back to session_index for the summary text.
• _merge_results() fuses keyword + vector hits with reciprocal rank fusion (k=60), so BM25 and cosine scores combine without normalisation.
• search() stays backward compatible: falls back to keyword-only when no embedding_fn is configured.
• FTS extension loading is now lazy and cached per process. Pure vector workloads and offline environments no longer pay the extension download cost, and _keyword_search() degrades to an empty result set (with a warning) when FTS can't be loaded — vector-only hybrid search still proceeds.

Cross-session loading (qracer/cli.py)

• qracer repl now creates ~/.qracer/summaries/, instantiates a file-backed MemorySearcher at ~/.qracer/memory_index.duckdb, indexes every Markdown summary under summaries/, and prints ✓ Loaded N past session summaries from … so returning users immediately see how much prior memory is in scope.
• Both memory_searcher and summaries_dir are passed into the ConversationEngine, closing the loop.

Tests

• tests/memory/test_memory_searcher.py — new TestHybridSearch class covering has_embeddings flag, embedding row storage, semantic vector search, hybrid search(), _merge_results() RRF ordering, removal from both tables, empty vector branch when no embedding_fn, and graceful handling when the embedding function raises.
• tests/conversation/test_engine.py — new TestCompactionPersistence class with two tests: (1) _maybe_compact() writes the Markdown file and auto-indexes the row when summaries_dir + memory_searcher are set; (2) without summaries_dir, it still calls the in-memory compact() path for backward compatibility.
• tests/conversation/test_topic_resolver.py — updated the existing FTS-availability skip pattern to pre-warm FTS now that MemorySearcher() construction no longer triggers it.
• docs/memory-system.md — removed the 구현 예정 markers and described the new wiring.

Results

• uv run pytest → 636 passed, 13 skipped (skipped tests require DuckDB's fts extension, which this sandbox can't download; they run on CI with internet access).
• uv run ruff check . → clean.
• uv run ruff format --check → clean.
• uv run pyright → 0 errors.

How to test

uv run pytest tests/memory/test_memory_searcher.py tests/memory/test_session_compactor.py tests/conversation/test_engine.py -v
uv run ruff check . && uv run ruff format --check . && uv run pyright

End-to-end smoke test:

uv run qracer repl
# have a conversation long enough to trigger compaction (> 8 000 tokens of log)
ls ~/.qracer/summaries/            # new markdown file
uv run qracer repl
# startup should print: ✓ Loaded 1 past session summaries from …