Preserve regression collector outcomes and canonical process context - #9234

Summary

Make the existing Python 3.15 regression-data collector diagnosable and give it the canonical test-process context:

• Preserve full test IDs and every subtest callback reported by unittest, including successful/skipped outcomes and failure/error tracebacks. Print error details at the existing verbosity.
• Record module/class fixture failures and skips without inventing unittest test executions. Count each completed parent/fixture record once, while retaining unittest's separate event-based summary.
• Reuse test.libregrtest.setup.setup_process() before discovery. Remove the collector's redundant RUSTPYTHONPATH override; the same-checkout interpreter discovers its compiled-in Lib normally.

No Lib changes, new workflow, matrix expansion, test skip, or assertion change to canonical tests. A subsequent coverage repair adds the established dev-profile VM optimization setting described below.

Validation

• Five focused infrastructure tests pass on pinned CPython 3.15.0 and a source-bound RustPython release: mixed/nested subtests, repeated IDs, interleaved suites, parent/cleanup errors, expected/unexpected successes, 12 module/class fixture skip/error variants, and interruption propagation.
• Canonical test_unicode_guard_env and test_dump_path_config pass in bounded owned children after canonical setup, on both runtimes without a path override. Normal/-E/-I RustPython provenance identifies its compiled-in checkout Lib.
• A paired startup observation confirms both runtimes can legitimately recover from an invalid PYTHONHOME when their respective path variable supplies encodings. This change does not alter runtime path precedence.
• Normal commit hooks pass. The initial runner/workflow-only changes required no new Rust build. The later dev-profile coverage repair has the fresh, narrowly scoped native validation described below; broader native verification remains historical.

The completed mega collector run still reported 42,754 cases, 15 failures, 180 errors, 4,716 skips and 291 expected failures. This PR does not make that suite green or rewrite historical results. Corrected JSON parent/fixture totals are not directly comparable to previously incomplete/overcounted totals; the raw unittest event summary remains distinct. The existing interleaved-suite timing limitation is unchanged.

The current regression website tolerates the additive JSON fields and renders status strings generically. Its pre-existing summary omits failure counts; that separate website behavior is unchanged.

Hosted collector follow-up

The first completed collector on this PR's main-based source reported 43,315 unittest executions, 12 failure events and 212 error events. Full IDs and tracebacks exposed two runner problems now repaired: successful subtest ID formatting could raise on huge integers or failing repr methods (three recorded reporter errors), and already-qualified split-package names were prefixed twice (64 import errors). The repair preserves normal full IDs, uses an explicit parent-ID fallback when formatting raises, and qualifies only bare discovery names. It does not change the integer conversion limit in production or suppress genuine import/test errors.

All five focused infrastructure tests pass on pinned CPython and the existing RustPython release, including huge-integer/raising-repr parameters across success/failure/error/skip and bare/qualified discovery with a genuine import failure. The test temporarily pins and restores its own integer limit. A subsequent Linux infrastructure assertion exposed forced-color output in CI; its two traceback text assertions now reuse the standard color-removal helper. All five tests also pass in paired plain, PYTHON_COLORS=1 and FORCE_COLOR=1 runs. Production output is unchanged. The full hosted collector has not yet run on this repair; its historical counts remain unchanged. The previous hosted run confirms unicode_guard_env passes; dump_path_config remains expected-failure on this main-based source because the separate startup fix has not landed.

Dev-profile coverage repair

The d9eed4a coverage job stopped progressing after an XML-RPC import raised RecursionError in the multiprocessing manager accepter thread. This follow-up carries the existing integration-tree [profile.dev.package.rustpython-vm] opt-level = 3 setting into this branch. It does not raise Python recursion limits, change native stack guards or thread stack sizes, disable debug assertions, or change test/release profiles.

Fresh bounded local validation: the same uninstrumented dev source failed cold xmlrpc.client imports on both the main thread and an owned helper thread before the override, and passed both after it, retaining recursion limit 256 and debug assertions. The matched native evaluator frame allocation fell from 237,544 bytes to 776 bytes plus saved registers. A build using the committed manifest produced the same binary as the configuration-only comparison; both import probes passed again. Two exact dev-profile native tests also passed: default_python_thread_stack_size_debug and explicit_python_thread_stack_size_is_a_floor_debug. Normal commit hooks passed.

Full workspace/CAPI tests, strict workspace/CAPI Clippy, multiprocessing/forkserver suites, and instrumented coverage were not rerun locally; the broader build gates exceeded the existing disk preflight. Prior broader native results remain historical evidence only. The matched imports establish the local dev-profile failure and repair, not that the complete hosted stall is resolved. Hosted instrumented coverage on this commit is still pending.

AI assistance

Prepared at the maintainer's request as part of the ongoing Python 3.15 upgrade. OpenAI Codex assisted with implementation, source review, diagnostics and automated validation. No claim is made that the bot account is a maintainer or that these checks were manually performed by a human. Human review and platform CI remain necessary.