parity-harnesslisted
Install: claude install-skill jaewhi-park/GARAGISTE
# parity-harness
## Principles
- No rebuild step starts without the harness.
- Bugs are recorded as current behaviour for now. Preserve-or-fix is the CEO's decision and goes in the decision log.
- Normalization rules (timestamps, ordering, floating-point tolerance, temporary IDs) are documented explicitly.
## Procedure
1. Collect representative inputs: real-data samples (sensitive data masked), boundary values, failure cases, one large case
2. Run the legacy system and store outputs as golden files — tests/golden/<case>/{input,expected}
3. Write a runner: select an implementation (legacy|new), feed the same inputs, diff against golden. Put normalization rules in the runner as code.
4. Validate the runner against legacy — everything must PASS for the harness to be valid.
5. Judge the new implementation with the same runner. Intentional differences: update the golden file and record the reason in docs/DECISIONS.md.
## Output
- tests/golden/ and the runner, registered under "Commands" in the rules file
- docs/PARITY.md: covered scenarios / known differences and reasons / uncovered areas