knowledge-mcp: Skill for Claude Code

CENSUS-AND-REUSE — enumerating surfaces and citing what already exists

Programmatic census

Any surface larger than ~15 sites or ~5 files, or pattern-defined, is enumerated PROGRAMMATICALLY (ast/grep/script, commands recorded, run during planning) — hand counts rot and do not converge under review. The census output IS the surface: per-file lists are floors; every sweep completion criterion RE-RUNS the census and asserts remainder-by-kind = 0. Multi-kind migrations get a small checked-in census script emitting a manifest ({file, line, kind}) with judgment sites marked manual. Pattern breadth: aliased forms, template literals, comment occurrences (state whether they count), indirect flows via callers — every kind the site definition matches needs a classification, or the gates are permanently unsatisfiable.

UNIFORM structural edits across many files are prescribed as ast operation:"replace" (dry-run preview, where-tree scoping, re-parse gate) — never sed/perl, never enumerate-then-Edit when one template covers every site. Sweep size is NOT an architecture constraint: never pick a lesser design to dodge a uniform sweep.

Counts are floors

Before any count or completeness claim becomes load-bearing, re-derive it with a pattern strictly BROADER than the one that produced it, then subtract. An exact-phrase sweep is a floor: one inflection, one casing, one quoting difference hides members. Helper indirection defeats literal-pattern censuses in three disguises: a helper taking the payload under another name, an anonymous inline struct, a synthetic internally-manufactured payload — after the literal pass, re-derive keyed on the CONSUMED TYPE and reconcile. A denominator you cannot rebuild row by row is an estimate wearing a measurement's clothes.

Caller census before touching a symbol