feat(skill-tree): learned procedural memory the maestro grows and retrieves (#458)

[hadamrd]

Implements the skill tree epic (#458): a living, evolving memory of how to do X in this repo that the loop writes from its own critic-clean merges and reads back into the next worker's brief — so workers stop re-deriving repo conventions cold.

It wakes forge-loop's dormant procedural-memory substrate (the SqliteMemoryStore + record_procedural_skill) rather than adding a parallel system.

The tree

Each card carries an area:<path> tag — its address in the tree. area_ancestors walks most-specific-leaf → ancestor node.

• Leaves = concrete procedure cards (trigger + recipe + pitfalls, stamped with the merged commit SHA).
• Internal nodes (area-node) = generalised patterns distilled once enough sibling leaves accrue.

Lifecycle

1. Harvest — a one-shot librarian (skill_librarian.py, LLM behind an injected seam) distils diff + acceptance → SkillCard; harvest_skills_from_merge records it with SHA provenance. Wired best-effort into the post-merge hook (runner/tick.py), gated by misc.skill_tree_harvest (FORGE_SKILL_TREE_HARVEST, on by default).
2. Retrieve + inject — area-weighted deterministic ranking (memory/skills.py) feeds render_skill_section into make_brief + make_repair_brief; emits skill_injected for measurement. Retrieval is cheap + local so it's always on.
3. Curate — expire_stale_skills retires cards whose proof SHA aged out of history; promote_internal_nodes generalises sibling leaves. Events: skill_harvested / injected / promoted / expired.

Design traps engineered around

• Stale > empty — every card carries provenance SHA + confidence; expiry drops stale cards from retrieval (adversarial test: an expired card matching the query is NOT returned).
• Retrieval is the hard part — honest deterministic area/keyword ranking now, with a clean seam for semantic (Lumen) retrieval as a documented fast-follow.
• Contradiction sprawl — one librarian writes; supersession preserves lineage; no free-for-all.

Tests (TDD, adversarial)

New: test_skill_area_tags, test_skill_librarian, test_skill_harvest, test_skill_events, test_skill_retrieval, test_skill_curation, test_worker_brief_skills — each behavior has a test that goes red if the harvest/inject/expiry is removed. Full suite green except pre-existing test_runner_retry.py cases that require a GitHub token (they fail identically on clean trunk in a token-less shell). Lint + format clean.

Out of scope (fast-follow)

Semantic/Lumen retrieval (seam only); a UI surface; cross-repo skill sharing.

Spec: docs/design/81-skill-tree.md.

🤖 Generated with Claude Code