Fix memory pipeline quality: ranking, LLM gating, decay, noise suppression

[CalebisGross]

Summary

Addresses the core pipeline issues identified by the system audit — the memory system was amplifying noise instead of filtering it.

• FTS5 BM25 ranking: SearchByFullText now uses ORDER BY rank instead of insertion order, so recall returns the most relevant matches first
• Dreaming agent LLM gate: Skip the entire dream cycle when LLM is unavailable — without it, dreaming was blindly strengthening associations on junk data
• Consolidation LLM gate: Skip merge and pattern extraction steps when LLM is down, but still run DB-only maintenance (decay, transitions, pruning, expiry)
• Pattern dedup threshold: Lowered from 0.8 to 0.70 cosine similarity to catch near-duplicate patterns that were slipping through
• Decay tuning: Reduced recency protection from 0.5x to 0.8x (24h window) and 0.8x to 0.9x (7d window) so low-quality memories fade in ~2 cycles instead of 4+
• Heuristic noise suppression: Added path-aware zero-scoring for application-internal directories (Chrome/Chromium, LM Studio, GNOME dconf, Trash, browser storage dirs)
• Hook bugfix: Fixed protect-git.sh false positive — regex matched -f in branch names like fix/audit-pipeline-fixes
• go fmt: Reformatted 23 files that had drifted from standard formatting

Test plan

• make build passes
• make check (go fmt + go vet) passes
• make test — all tests pass, including updated decay test expectations
• Verify with live daemon: nuke DB, restart, confirm Chrome/LM Studio events get 0.0 score in logs
• Verify dream cycle logs "skipping dream cycle: LLM unavailable" when LM Studio is stopped
• Verify FTS recall returns more relevant results on real queries

🤖 Generated with Claude Code