Defects and Evidence

One ledger

Defects found by /test-run go into syntaxis/context/findings.md — the same findings ledger that /audit writes to.

A failed test therefore blocks /complete through the gate that already exists. No second gate, and no way for code review and testing to disagree about whether work can ship.

### F-07 [P1] open - Checkout accepts an expired card

**File:** src/checkout/validate.ts:42
**Found:** 2026-09-19 by /test-run (cycle: regression; case: TC-014)
**Why it matters:** ...
**Suggested fix:** ...
**Resolution:**

Rules:

• IDs continue the ledger’s existing sequence, never reused, never renumbered.
• Severity is P0–P3 by the consequence of the defect, not by how visible it is. Only P0 and P1 block completion.
• File names the suspected location when the cause is identified. When it is not, the record says the cause is unlocated rather than guessing a file.
• One finding per distinct defect. Several cases failing from one cause produce one finding listing those cases, not one finding each.
• A blocked case is not a defect. It goes in the cycle record only, unless the thing blocking it is itself a defect.
• /test-run never changes an existing finding’s status. Re-verifying a repair is a later cycle’s job; closing a finding belongs to /audit.

The cycle record

Every run writes syntaxis/qa/runs/<YYYY-MM-DD>-<cycle>.md:

# Test run - regression - 2026-09-19

**Branch:** feature/checkout-limits  **Commit:** a1b2c3d
**Environment:** local  **Build:** 0.4.2
**Platforms:** Chrome 141 (macOS 26), iPhone 17 simulator (iOS 26)

## Summary
| Result | Count |
| pass | 18 |
| fail | 2 |
| blocked | 1 |

## Results
| Case | Result | Platform | Finding |
| TC-014 | fail | web | F-07 |
| TC-021 | pass | iOS | |
| TC-022 | blocked | iOS | |

## Failures
### TC-014 - Checkout rejects an expired card
**Observed:** The order was accepted and a charge was created.
**Expected:** The order is refused and no charge is created.
**Steps to reproduce:** ...
**Evidence:** test-results/tc-014-failure.png
**Finding:** F-07

## Exit criteria
Each criterion from the test plan, and whether it was met.

The commit, environment, and platform versions are mandatory. A result without them proves nothing a week later.

Rules that keep a cycle honest

Rule	Why
Never edit application source during a cycle	Repairing mid-cycle invalidates the record
Never weaken or rewrite a case to produce a pass	A wrong case is a case defect, fixed separately
Observe, do not infer	A case passes because its result was seen, not because the code looks right
A partial pass is a fail	And the record names which half failed
A flaky case is a fail	Until proven otherwise; the attempts get recorded
Blocked is not a soft fail	Use it only when the case truly could not run, and name the blocker

Evidence for every failure: exact reproduction steps, observed versus expected, screenshots or recordings the harness produced, and relevant logs — with credentials, tokens, and personal data redacted.

Entry criteria

A cycle refuses to start against an unknown build. /test-run records the branch, exact HEAD, environment, and platform versions first, and checks the test plan’s entry criteria. Running against an environment the plan excludes produces a record that means nothing.

After a cycle

Situation	Next
A failure with an unclear cause	/debug
A failure with a known cause	/fix
Exit criteria genuinely met	/complete

/complete archives the cycle records for the work item under syntaxis/history/test-runs/. Records are immutable: a superseded cycle is never edited or deleted, because it is evidence of what was true at that commit.

What a report always says

The last line of every cycle is what it did not cover. A cycle that ran on one browser or one simulator proves only that, and the report says so rather than letting a summary count imply more.