great_cto

Ranked memory recall (BM25)

• scripts/lib/memory-search.mjs — a zero-dep BM25 search over great_cto's
  memory (session logs, brain/lessons/decisions, and ~/.great_cto/global-patterns).
  Rich memory was let down by naive recall: /resume read the most recent
  logs and architect's pattern-lookup grepped patterns by exact applies_to
  frontmatter. Recent ≠ relevant, and an exact grep misses a related pattern
  under a different label. Ranked search surfaces the relevant memory.
• Wired into /resume — now also surfaces the past sessions most relevant
  to the current work (query built from the git branch + open task titles), not
  only the newest three.
• Wired into architect pattern-lookup — adds a "also possibly relevant
  (ranked recall)" section alongside the exact-tag matches, so a semantically
  related pattern surfaces even when its applies_to label differs.
• Pure Node, no deps, no index files, builds in-memory per call (memory corpora
  are small). Fail-open: silent if node/module is unavailable. Idea adapted from
  Kuberwastaken/reference (BM25 over agent memory); implemented fresh in Node.
• Tests: lib suite +6 (348 total, all green).