fix(skill): narrow implicit codebase routing

[tadeogutierrez]

Summary

• narrow Graphify's implicit skill description to relational and cross-file analysis
• prefer direct source search for symbol/literal lookup, single-file inspection, and localized edits
• preserve explicit /graphify behavior across all generated host variants

Evidence

Six clean Hermes canaries loaded Graphify for every codebase task under the current description, including symbol lookup, literal lookup, and single-file inspection. No graph query ran because the test repository had no graphify-out/; the cost was skill activation and context loading rather than useful graph retrieval.

Test plan

• uv run --with pytest pytest tests/test_skillgen.py -q — 65 passed
• uv run ruff check tests/test_skillgen.py
• python -m tools.skillgen --check
• python -m tools.skillgen --audit-coverage
• python -m tools.skillgen --schema-singleton
• python -m tools.skillgen --monolith-roundtrip
• python -m tools.skillgen --always-on-roundtrip
• git diff --check
• sabotage check: the updated descriptor regression test fails against unmodified v8

Scope

This intentionally changes only implicit routing. It does not add a freshness/provenance contract; that broader work needs to account for AST vs semantic hashes, persisted corpus configuration, and monolith parity.

Closes #2216

[graphify-labs]

Graphify reviewed this change.

Looks safe to merge — no coupling regressions and no blocking issues, checked against the code graph (not a self-assessment).

---

Graphify review — findings

Rewrites the graphify skill description frontmatter across all agent variants to scope invocation to explicit Graphify requests, relational queries answerable from an existing graph, and cross-file architecture/dependency/impact analysis, while steering simple symbol lookups, single-file inspection, and localized edits toward direct code search. Updates the skillgen fragment sources, expected-output fixtures, and rationale/enum tests to match the new wording.

No blocking issues surfaced. 7 lower-confidence candidates did not survive cross-model review.

Analysis details — impact, health, verification

Impact & health

Graphify review

Impact — 1034 functions depend on the 1034 functions this change touches.

Health — this change adds coupling hotspots:

• new: test_audit_catches_a_dropped_non_allowlisted_heading() — 0 callers, 6 callees

Verification — 1034 functions in the blast radius were not formally verified this run (proofs are advisory here).

Gate & verification

graphify gate

PASS — objectively clean (no health regressions, tests not run — proofs not run this pass (advisory)). Grounded, not self-assessed.

Advisory (not blocking):

• verification_scope: 1034 function(s) in the blast radius were not formally verified this run

· 1 more finding(s) on lines outside this diff (see the check run).