fix: make the schema codegen host-independent

[PerryLink]

The defect

The schema code generator's output depends on the host OS, in two places. The generated modules are checked in and make gen-check (also exercised by tests/test_gen_all.py::test_codegen_check_is_clean_and_read_only) asserts that regenerating them reproduces the committed artifact — so a generator that is not a pure function of its inputs defeats the check it is there to satisfy.

1. The header embeds os.sep. scripts/gen_schema.py:179 and scripts/gen_meta.py:59 interpolate a Path into an f-string, which stringifies with the platform separator. The committed header says schema/schema.json; regenerating on a non-POSIX host produces:

-# Generated from schema/schema.json. Do not edit by hand.
+# Generated from schema\\schema.json. Do not edit by hand.

2. The ruff pipe is decoded with the ambient locale. Both generators run ruff through subprocess.run(..., text=True) without encoding. text=True uses locale.getpreferredencoding(False), so the child's UTF-8 output is decoded as whatever the host locale says — on this machine cp936:

locale.getpreferredencoding(False) == 'cp936'

When ruff exits non-zero, the decode fails before the message is read, and the diagnostic is replaced:

RuntimeError: ruff check --fix failed:
ruff failed
  Cause: stream did not contain valid UTF-8

The real reason ruff failed is discarded. That is the error path of a build script, so the loss is the whole point of the message. This also hides defect 1: the decode error is raised before the artifact comparison, so the header mismatch is never reached.

The change

Four lines, no behaviour change on a host whose locale is already UTF-8:

-    lines = [f"# Generated from {schema_json.relative_to(ROOT)}. Do not edit by hand."]
+    lines = [f"# Generated from {schema_json.relative_to(ROOT).as_posix()}. Do not edit by hand."]

             input=source,
             text=True,
+            encoding="utf-8",
             capture_output=True,

encoding="utf-8" is what the rest of both files already does for every file they read or write (gen_schema.py:64,81,91,180, gen_meta.py:26,43), so this makes the subprocess consistent with its own module rather than introducing a convention. as_posix() emits / on every host, which is what the committed artifact already contains.

Verification

tests/test_gen_all.py is the repository's own check for exactly this, and it needs no new test:

                        before            after
tests/test_gen_all.py   1 failed, 20 passed    21 passed

Reverting only the two scripts and keeping everything else restores 1 failed, 20 passed, so the change is what the test detects rather than an incidental pass.

The rest of the suite is unchanged by this: --doctest-modules goes from 11 failed / 346 passed to 10 failed / 347 passed, the one difference being this test. ruff check and ruff format --check are clean on both files.

Scope

I did not touch the remaining failures in this checkout. They are in tests/http/* and tests/real_user/test_stdio_limits.py and I have not established whether they are host-specific or real, so I am not claiming anything about them. CI runs on ubuntu-latest only, where the header separator is already / and the locale is already UTF-8 — which is why neither defect is visible there, and why this is a reproducibility fix rather than a platform feature.