[bvolpato]

Description

Motivation. A parallel unittest process can exit unsuccessfully after its tests pass. With an empty last-failed cache, the wrapper's --lf retry executes the entire passing workload again. A successful rerun can then hide the process failure. This wastes test execution and makes the CI result unreliable.

Change. Give each parallel attempt and its serial retry a private pytest cache, and set --lfnf=none. Pytest teardown removes the cache.

Verified before/after. The reproduction injects a nonzero controller exit after a passing run with real pytest-xdist workers:

Scenario	Before	After
Process fails without a recorded failed test	Passing workload runs again; wrapper can return success	Passing workload is not rerun; wrapper remains failed
A test fails once and passes on retry	Retry recovers	Targeted retry still recovers

The regressions demonstrate failure preservation and avoided re-execution. Saved CI minutes and changes in failure frequency have not been measured. The original process failure still requires diagnosis.

Test Coverage

• Extended test_unittest_culprits.py with actual pytest-xdist worker processes and actual sequential last-failed selection. Empty and stale caches preserve process failures, while a test that fails once can pass its targeted retry.
• The regressions fail against the original wrapper and pass with the change on Python 3.12 with pytest 8.4.2 and 9.0.3.
• The helper module is marked cpu_only and is covered by the existing unittest/tools CPU CI entry.
• Scoped pre-commit and git diff --check pass. Jenkins and GPU workloads were not run locally.

PR Checklist

Please review the following before submitting your PR:

• PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

• PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

• Test cases are provided for new code paths (see test instructions)

• If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• Update tava architecture diagram if there is a significant design change in PR.

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.