Guard the tools that guard this repo: fail CI on a name nothing binds

[andrewboudreau]

Guards the tools that guard this repo

#1392 fixed a NameError that had blocked PR validation outright. The module was
bound as AP and read as asm_policy, on a branch only reachable when a file graded
NO-REPRO — so the first PR to contain a declared draft got a traceback instead of a
verdict, and the worker discarded the results it had already computed.

A gate's failure branch is the one path a green CI run never exercises. No ROM
build will ever catch a defect there — these tools are not compiled into the ROM — and
only static analysis reaches that code without running it. This is the job that stops
it happening twice. It finds the #1367 defect in under a second across all 186 tracked
Python files.

Narrow on purpose — this is the part worth arguing with

The gate fails on unresolvable names and unparseable files, and nothing else.
The tree has zero of those, so this lands green and every future red is a real
regression.

It also has 32 style findings — unused imports, unused locals, f-strings without
placeholders. Gating on those meant one of two bad outcomes: landing this red, or a
32-file cleanup commit in the same PR that buries the actual signal. A gate that lands
red gets switched off, and then it catches nothing at all.

So the style findings are counted on every run, listed under --advisories, and not
gated. Ratcheting them downward the way langmode-ratchet.yml does is a fair
follow-up — it needs a banked baseline, and this deliberately does not.

pyflakes is pinned to 3.0.1. A floating version can add a check, and this job
turning red on a PR that touched no Python would be indistinguishable from a real
regression — exactly the confusion a stale ratchet caused earlier this week.

Verified

• The One banner rule instead of three, and let the PR gate fail a file that won't build #1367 defect, reconstructed verbatim, is caught:
  repro.py:5: undefined name 'asm_policy' → exit 1.
• An unparseable file is caught, and reported separately from the name check.
• 186 tracked files: 0 unresolvable, 0 unparseable, 32 advisory.
• 7 tests pass under python -m unittest tools.test_check_python_names, which is how
  the workflow runs them — stdlib only, so the runner needs no dependency beyond the
  pinned pyflakes.
• Adds no src/ or config/ change, so the ROM build is untouched.

The gate caught its own author on the way in: the first draft imported
pyflakes.reporter without using it and pushed the advisory count 32 → 33.

Run it locally

python tools/check_python_names.py               # every tracked *.py
python tools/check_python_names.py --advisories  # also list the style findings

🤖 Generated with Claude Code

https://claude.ai/code/session_01XxDmkQ47fWa3GB6mj8xEQe

[tangos-validator]

✅ PR validation — Passed

noverify: no source/build-data changes in this PR

_{Each changed src/*.c|*.cpp is compiled and its relocated bytes compared to the binary data on a private build box. Passing requires every changed file to reproduce the ROM byte-for-byte with correct relocation targets — this catches WRONG-DEST relocations and non-reproducing near-misses that ledger-scoped linkcheck skips.}