[codex] Preserve seeded file-order reproducibility when randomizing test runs

[bpc-oss]

Summary

This fixes a reproducibility bug in seeded test-file randomization.

When --randomize or --seed is used, file ordering should be determined by the explicit file set plus the seed. Before this change, AVA still applied failingTestsFirst() before shuffling files. Since that cache-based reordering changes the input array, the same seed could produce a different file order after previous failures updated node_modules/.cache/ava/failing-tests.json.

This patch:

• skips failingTestsFirst() when randomization is active
• adds an integration test proving the same seed yields the same file order even when the failed-test cache changes

Why

Issue #595 explicitly asks for randomization that remains reproducible for debugging. Hidden cache state should not change the order produced by a given seed.

Validation

• npx xo lib/api.js test-tap/integration/assorted.js
• $env:CI='1'; npx tap test-tap/integration/assorted.js
• npx tap test-tap/runner.js
• repo-native Linux rerun in node:24.15.0-bookworm with the repo extracted at /tmp/ava/ava:
  • npx tap test-tap/reporters/default.js
  • ./scripts/test.sh
  • ./scripts/ci.sh

Notes

• This intentionally preserves reproducibility over failed-test prioritization while randomization is enabled.
• Existing non-randomized runs continue to use failingTestsFirst().
• Earlier local failures were environment false negatives caused by Windows checkout issues and a temp path that prevented reporter stack-frame filtering. The fresh Linux rerun above is the authoritative validation evidence for this patch state.