feat: add FTS5 full-text search

[valiantone]

What

Add SQLite FTS5-backed keyword retrieval and use BM25 as the lexical component in hybrid ranking.

Changes:

• Creates a memories_fts FTS5 virtual table backed by memories.fact_text.
• Adds INSERT/UPDATE/DELETE triggers to keep the FTS index synchronized.
• Rebuilds the FTS index on DB init so existing databases get populated.
• Adds MemoryDB.fts_search(query) with raw BM25 scores.
• Replaces naive whitespace keyword overlap in search_memories() with normalized BM25 scores.
• Uses the porter unicode61 tokenizer and prefix query terms to improve stemming/prefix behavior.
• Applies TTL filtering to FTS results after feat: add per-memory TTL expiry #21 so expired memories do not contribute lexical scores.
• Adds coverage for FTS search, replace sync, hybrid ranking behavior, and the combined TTL/FTS path.

Closes #2

Dependency

Resolved: this branch has been rebased onto main after #21 merged, so FTS now includes the TTL schema and filters.

Why

The old keyword score only split on whitespace and counted overlap. FTS5 gives HotMem a credible local lexical retrieval layer while keeping the existing API response shape and SQLite-only architecture.

Verification

• Tests pass: PYTHONPATH=src /Users/zubinj/forge/.knwb/bin/pytest -> 42 passed in 2.01s on Python 3.13.5
• Lint passes: /Users/zubinj/forge/.knwb/bin/ruff check src/ tests/ -> All checks passed
• Format check passes: /Users/zubinj/forge/.knwb/bin/ruff format --check src/ tests/ -> 18 files already formatted
• No new external dependencies added to core