[maceopark]

What and why

main's checks have been red since the shader builder's instruction scan (#1036), and every pull request based on it inherits the failures: ruff and pyright (ratchet) reports two findings, and Tool tests fails two test files. Neither comes from a branch's own change.

Lint. Ruff B023 at dm/games/settings.py:115: offered() closes over the loop's row number n, so the row is now a default argument. Pyright's reportOptionalIterable in dm/tools/tests/test_split_linkdata.py: linkdata_jobs() returns None for a file it cannot read, so the test narrows the result before iterating it. This is the same fix as #1235, which covers the first finding alone; if #1235 lands first, this branch rebases to the two test files.

test_dmsh_pool_arrays. The test loads dmsh_build.py by executing only the functions it names. reflect() gained a call to has_word(), which calls in_word(), and neither was in the list, so reflect() raised NameError: name 'has_word' is not defined. load() now pulls in every function the named ones call, so the next helper a tested function gains comes with it.

test_image_atomic_bounds. The test pinned uses_image_atomics() raising on a zero-length instruction in a module that holds no OpImageTexelPointer. Since the instruction scan, a module without that word is answered "no image atomics" without being walked, so the malformed check runs only where the walk does. The test now pins both halves: the walk's ValueError, on a module with the opcode's word before the malformed instruction, and the fast path's False for the module without it.

Making the malformed check exact again would put a scan for a zero instruction word in front of the walk, in dmsh_atomic.py, which is a chain file: the new chain version would rebuild every game's shader library to change nothing on a well-formed module. So the fast path's contract is pinned instead. That is open if a malformed module's error message matters. Nothing here changes what the chain computes, so dmsh_chain.h is untouched.

Checked

• Tests: python3 -m unittest dm.tools.tests.test_dmsh_pool_arrays dm.tools.tests.test_image_atomic_bounds dm.tools.tests.test_split_linkdata (21 tests, OK, 7 skipped locally: the Metal compiles CI runs), and dm/tools/tests/test_settings*.py (19 tests, OK).
• dm/tools/lint_python.py --against origin/main with the pinned tools from .github/requirements/lint.txt: 389 findings, 389 in the baseline, 0 new.
• Not run: scripts/regress.py, which reports nothing for a chain that does not change; no game was run.

Docs

• No reference entry: the chain's outputs do not change on a well-formed module, and the contract the two tests pin is now written in the tests.
• No chain tool changed: dmsh_chain.py --header was not run.
• No game files, keys or data extracted from a game in the change.

[seedds]

Thanks for the lint and fixture repairs. The remaining two-file diff is now covered on main by 58a77cc2: pack_quiet has the days: float annotation, and the NSP test patches NSP_FOLDER with mock.patch.object instead of assigning the module global in its callbacks.

Closing as superseded by the current implementation; no additional changes from this remaining diff need to land.