# Impacted regression-test policy

`impacted_tests.json` is the reviewed changed-path map used by `fold quality`.
Every run retains `test-selection.json` with the changed paths, matched rule IDs,
selected tests, and rationale for each selection.

Two suites are unconditional:

- `fold:benchmark_evaluation` preserves positive, negative, and regression
  corpus contracts.
- `fold:source_structure` preserves architecture and source-budget contracts.

Rules add subsystem-specific integration and regression tests. Selection is a
union, so a change spanning subsystems runs every applicable test. A changed
production, build, tool, or test path covered by `fallback_globs` but by no
reviewed rule runs the complete Meson suite. A missing or invalid manifest and
git-diff failures also fail closed to the complete suite.

When introducing a subsystem, update the manifest in the same change. Name the
behavioral tests that exercise its public effects and give a rationale. Do not
remove unconditional or fallback coverage merely to reduce CI time; first add
evidence that a narrower rule preserves the relevant regression coverage.

## Flake evidence

`fold quality --test-repetitions N` executes the same selected set independently
`N` times. Each attempt retains its complete console log and a sanitized result
artifact. `flaky-test-evidence.json` classifies each test as `stable-pass`,
`stable-fail`, `stable-skip`, `flaky` (both pass and fail), or `inconsistent`
(missing/skip transitions or malformed evidence). Sanitized artifacts omit the
Meson environment and test stdout because the complete console log is retained
separately and environments may contain credentials.

A later pass never erases an earlier failure: every attempt and the aggregate
classification participate in the gate status. Repetition is evidence, not an
automatic retry or quarantine mechanism. The default is one attempt; broader CI
tiers may request repeated evidence explicitly.

## CI tiers

`.github/workflows/quality-gates.yml` applies one explicit tier per event:

- **PR (pull request):** merge-base differential source scan, impacted
  regressions once, and the Fold policy manifest.
- **Nightly (daily schedule or manual dispatch):** complete Meson suite three
  times plus the Fold policy manifest, with no branch differential.
- **Release (published release or manual dispatch):** complete Meson suite,
  Fold policy manifest, reviewed 70% line-coverage floor, and release-binary
  smoke.

Every tier uploads `quality-report/`, including complete logs, SARIF where
applicable, test selection, repeated-test classifications, policy summaries,
and compact token measurements. The nightly repetition count collects flake
evidence; it does not convert failures into passes. Release coverage uses the
previously reviewed floor rather than introducing a new threshold.