Ratchet backlog: 229 file + 607 method size-cap breaches pre-date the new craftsmanship-checker gate (issue #297)

[cdeust]

Context

Issue #297 wires a real, blocking CI gate (scripts/craftsmanship_checker.py)
enforcing CLAUDE.md's documented-but-previously-unenforced 300-line-file /
40-line-method (signature-to-close) caps. Running it against the full tree
at commit 2590922 (origin/main, 2026-07-30) found a large pre-existing
backlog, measured via git ls-files -- '*.py' + Python ast:

Scanned 1093 tracked .py files, 12109 functions/methods.
FILE breaches (>300 lines):   229
  by top-level dir: mcp_server 121/527, tests_py 74/459, benchmarks 25/63,
  scripts 8/39, video 1/1, fuzz 0/3, _pipeline 0/1
METHOD breaches (>40 lines, signature-to-close): 607
  by top-level dir: mcp_server 401/3424, benchmarks 74/436, tests_py 85/7937,
  scripts 39/235, video 6/20, fuzz 0/11, _pipeline 0/1

Fixing 836 breaches spanning nearly every module in the repo is a
repo-wide refactor an order of magnitude larger than "wire the gate"
(coding-standards.md §15.1: don't mix an unrelated, much-larger root-cause
fix into the gate PR) and would collide with the many other in-flight PRs
already touching these same files.

What issue #297 does instead

Ships a committed, dated ratchet baseline
(scripts/craftsmanship_baseline.json) recording exactly this backlog as
of 2590922. The checker:

• Fails any NEW breach (a file/function not in the baseline going over cap).
• Fails any REGRESSION (a baselined item getting even one line longer).
• Fails a STALE baseline entry (an item fixed or removed must have its
  baseline entry removed in the SAME change that fixes it) — this is what
  makes the ratchet monotonically shrink over time instead of accumulating
  dead entries.
• Passes a baselined item that stays the same or shrinks (still over cap).

Acceptance criteria for THIS issue (the backlog, not the gate)

• Reduce scripts/craftsmanship_baseline.json's files and methods
  entries to zero, one PR (or a small number of scoped PRs) at a time,
  each a behavior-preserving Extract-Function/Extract-Module refactor with
  the existing test suite as the proof (passing unchanged) plus new tests
  for any newly-extracted unit, per coding-standards.md §6/§9.
• Highest-value first: the 20+ largest offenders
  (mcp_server/infrastructure/pg_schema.py at 2482 lines,
  mcp_server/core/wiki_coverage.py at 1444, mcp_server/core/ recall_pipeline.py at 1386, mcp_server/infrastructure/pg_store.py at
  1380, etc. — full list in the checker's --report output) are both the
  biggest readability cost and the largest single chunks of the ratchet.
• Each landing PR re-runs python scripts/craftsmanship_checker.py --update-baseline and commits the shrunk baseline as part of that PR
  (the checker enforces this: a fixed item left in the baseline fails as
  stale).

Not blocking any other work; this is the backlog issue #297's completion
ledger explicitly defers to, per coding-standards.md §13.4/§14.3.

[cdeust]

Closing in favour of correcting the false claim rather than tracking an 836-item backlog.

The premise of this issue is that CLAUDE.md advertises a 300-line/40-line cap 'Enforced by the craftsmanship-checker pre-commit hook'. That hook does not exist (no file matching craftsmanship, no .git/hooks entry, no .pre-commit-config.yaml, no CI job measuring method length) — so the caps have never been enforced, and 229 file + 607 method breaches accumulated behind a claim of enforcement.

Building a gate that fails on 836 pre-existing breaches would either block every PR or need a ratchet baseline that is itself a large piece of work. Neither serves the repo right now. The honest, small fix is to stop claiming enforcement that does not happen — handled in #297.