v0.2.7 — Raw extraction, lock recovery, hook hardening

Features

• Raw conversation filter — Replaced Haiku LLM extraction with deterministic raw conversation filter (a7e98373b). Faster, cheaper, no API dependency.
• MemoryProbe v2 — 5-category weighted scoring with 14 traps for benchmark evaluation (5a176244f)
• SYN-07 benchmark results — Discovered fundamental benchmark design flaw: Opus 4.6 implements Developer Notes into code, making "code-invisible" probes visible (ec4c5fcd4)

Bug Fixes

• Stale lock recovery — wt-memory run_with_lock() now detects and recovers from stale/orphaned lock files (21147e0c9)
• Hook config downgrade — wt-deploy-hooks can now downgrade stale hook configurations instead of failing silently (e9adca241)
• Usage bar gradient — Fixed gradient showing half the actual percentage (b3943ca7c)

Improvements

• Bulk archived 7 completed changes, synced all delta specs
• Gitignore for shodh-memory data in benchmark/synthetic
• Benchmark v7 update with metrics integration and recall-verify emphasis

Full Changelog

v0.2.6...v0.2.7