Complete PyTime conversion fixtures and repair existing helpers - #9317

Summary

Complete the coherent 11-helper PyTime conversion family in _testinternalcapi:

• Repair the three existing helpers: _PyTime_AsTimeval_clamp, _PyTime_AsTimespec, and _PyTime_AsTimespec_clamp. Honor all four rounding modes, negative values, ties-to-even, carries, native bounds, and CPython wrapper argument validation/precedence. Derive SIZEOF_TIME_T from the host type.
• Add the eight real conversion fixtures needed by the canonical PyTime tests: _PyTime_FromSeconds, _PyTime_FromSecondsObject, _PyTime_AsTimeval, _PyTime_AsMilliseconds, _PyTime_AsMicroseconds, _PyTime_ObjectToTime_t, _PyTime_ObjectToTimeval, and _PyTime_ObjectToTimespec.
• Keep the consumed numeric conversions in a private helper module. Windows timeval seconds use C long independently of 64-bit time_t; clamping clears the subsecond field.

This intentionally extracts the PyTime fixture/API prerequisite from the broader work in #9143 and #8954. The original three-helper repair and the additional eight fixtures are kept together in this PR rather than publishing a duplicate dependent PR. When #9143 is rebased, preserve or reuse these contracts instead of replacing them with less strict wrappers.

No _testcapi namespace, unused shared exports, time.sleep refactor, lifecycle changes, or Lib edits are introduced. These are Python-facing fixture implementations, not a claim of C-ABI coverage.

Validation

Pinned reference: CPython 3.15.0 final, d220409a13b77de7a6c30abe16ea90db01528edc. Oracle interpreter, libpython and _testinternalcapi extension hashes are bound to the retained build/source provenance.

Validated final commit: 29404850a370aad85467a782b65d7b44349a5f62.

• 2017/2017 fixed builtin-only observations for the eight added fixtures match CPython exactly, including typed results, exception types/messages, all rounding modes, signed ties/carries, integer boundaries, exact float inputs, arity/kwargs, invalid builtin modes/types, and conversion precedence. No unavailable helpers.
• 287/287 existing-helper regression observations match CPython. The original baseline had 99 differences (70 tuple values, 7 exception types, 22 messages); all remain resolved with no new differences.
• Normal repository commit hooks pass.
• CLI build with the root capi feature passes.
• Strict VM Clippy passes with default features and with threading (-D warnings).
• VM --no-default-features check passes, with 18 warnings in unchanged code; this is not a warning-free gate.

The original baseline binary was main 0b0a3607 plus the AST-Single warning patch from #9303. Its entire _testinternalcapi.rs matched the PR's main 94f92e4a base, and the exec-mode integer matrix does not exercise that warning route. This is not a claim of exact-current-main baseline or fresh-main integration testing.

Local runtime validation is Linux 64-bit. Integer 32-bit time_t probes and source-reviewed platform branches do not establish Windows/32-bit execution; float neighbors around ±2**31 are not covered. Custom conversion callbacks and subclasses are not exercised by these finite matrices.

Canonical PyTime classes require _testcapi even for the internal helpers. Their guards are unchanged, and skipped classes are not counted as passing coverage. No callback, sleep, subprocess, or lifecycle tests are claimed.

AI assistance was used in preparing this change and its validation.

Summary by CodeRabbit

• Bug Fixes
  • Time conversion test helpers now apply the requested rounding mode and report errors for invalid arguments, unsupported rounding values, and out-of-range inputs.
  • Clamped conversions return the nearest supported time bound with a zero fractional component.
• Compatibility
  • Time-size reporting and conversion limits now reflect platform-specific time types and bounds.

Rebase validation (2026-10-24)

Rebased onto main 1cd1b5d09f4853b2f13a5f3d5d6f856e901b671e as 4aee9e991297300a664593293d05bdb37316acdb. Both original commits are unchanged in git range-diff, and the full-index patch is byte-identical. No Lib, test, sleep, or lifecycle changes were added.

• Normal commit hooks, threaded VM cargo check, strict threaded VM Clippy (-D warnings), and the default-feature CLI build passed on the exact rebased tree.
• The 12 canonical numeric methods were selected with their existing imports and decorators: all 12 skipped; zero test bodies executed. This supplies compile/lint/build coverage, not numeric behavior validation. Earlier runtime observations above apply to the original commit, not a rerun on this rebased head.
• No full workspace, C-API, release, lifecycle, or cross-platform execution is claimed for this local rebase validation. Correctness CI run 40042903402 completed successfully on this exact rebased head: 25/25 jobs passed.