story-041: The verification runner is not assumed to be Python

[jerodw]

The coordinator executed Python source to record which interpreter ran the verification suite, and read a configuration key whose name was a language. A target repository whose suite is not Python got an empty record and was told nothing.

What changed

The version probe is gone. _VERSION_PROBE = "import platform; print(platform.python_version())", run as [interpreter, "-c", probe], was Python source and Python's flag for it. The recorded version field went with it: the interpreter path was already recorded and already answers "which one ran", so the field was not worth a new configuration key to keep.

clean_clone_python is now verification_runner — naming the role rather than the language. Nine coordinator sites, two artifact schemas, the config and the template.

A retired key is refused, not ignored. A config still carrying clean_clone_python would otherwise fall through to the replacement's default and quietly exercise something other than what it asked for. The refusal sits above every other pre-flight, because it is decidable the moment the config loads: no run directory, no state.json, no log, no branch, no agent invoked.

The burn-down

story-040 split the scan's allowlist into a temporary half that should empty and a permanent half that should not. This story is the first payment:

TEMPORARY_TIES   9 entries  ->  3

What remains is exactly the other story's subject — two lines of prose in prompts/tester.md and may_not_create: ["tests/"] in the workflow. the-test-location-comes-from-configuration removes those three, and the list emptying is story-040's completion signal for the whole audit.

Three judgements worth a reviewer's eye

One literal survives, deliberately. orchestration/harness_config.py names clean_clone_python in order to refuse it, which is nominally against this story's own constraint that no language name enters orchestration/. Task 6 required the refusal and a retired key cannot be refused without being named, so it is on PERMANENT_MENTIONS with that reason recorded. Filed as an-undeclared-config-key-is-refused — refusing any undeclared key removes the literal and catches mistyped keys, which nothing catches today.

One acceptance criterion was over-stated. AC10 expected the scan to report nothing in the touched files; one line remains — a self_route_problems docstring saying "True is not a budget however much Python is willing to treat it as one". That is a statement about this harness's own implementation, so it moved to PERMANENT_MENTIONS rather than being edited away.

One edit outside scope.modify, and it was forced. tests/test_stage_baseline.py compared story-037's pre-story bound against the working tree, so renaming a local in run_clean_clone made it red for a change story-037 has nothing to say about. Repointed, not weakened.

Verification

• 2288 passed, 21 written; verification passed on the first iteration, no retries; clean-clone green with the story committed
• Six implementer edits under tests/ were permitted by the revert check — the suite fails with them reverted

🤖 Generated with Claude Code