Verification, test execution, and command output

First read command output. Model: verification.

Testing

Never put literal PREFIX-N item ids in tests; use variables or pattern matchers.
Iterate with yoke watch pytest --impacted main --bounded. In a lane it first brings the lane current with origin; uncommitted work or a conflict refuses with the reason and fix and runs no tests. With ci_workflow_file it runs on CI, so commit first. --local is only for a check that ends in about one minute; past that, interrupt it, keep the capture as incomplete, commit and run on CI.
selection unbounded means keep testing what you judge relevant; the final gate covers the rest. --local --impacted main --widen is the CI-outage fallback only.
Run selections before the QA case. A selection follows a running or green full CI run on its commit and refuses a near-full selection, naming the QA case command.
Run the full suite once, through the attached case: yoke qa requirement list --item PREFIX-N, then yoke qa case run --requirement-id <id>. Commit before it; workers never push by hand. Never run the full suite twice on the same tree; rerun only after changes.
A quiet or backgrounded gate is still running: continue it. If its watcher was killed, rerun yoke qa case run; it rejoins or adopts the CI run for that commit.
An item case bound to an item-scoped deployment stage runs with yoke qa plan run --item PREFIX-N --stage STAGE. An attached post-deploy plan runs once per release as a plan group: yoke qa plan run --deployment-run-id RUN --stage item-qa --plan PLAN.
Never run raw python3 -m pytest over whole test roots; use yoke watch pytest.
CI verdict wake-ups need a healthy relay on this machine (CI wake-ups).
Several lanes: yoke qa case run --requirement-id N --checkout-path PATH --expected-branch BRANCH --expected-sha FULL_SHA. Follow refusal recovery for qa_source_lane_unresolved, qa_ci_source_mismatch and qa_published_head_unproven.

Unexplained test failure

The item holder may retry once if the change cannot explain the failed test. First message yoke say --steering with the test and CI run URL. A second failure of the same check on the same inputs belongs to this item. Failed trains use the recorded --retry-flake path.

QA results

Judge a result against the item's user outcome first. A case step the product cannot meet while the outcome holds is a case defect: repair the case through its registered write and rerun. A pass while the outcome fails is a product defect.
After failed QA: product defect, return to implementation; setup failure, record the blocker; unclear cause, investigate now.
A plan may record another environment, but standalone QA runs only where the runtime allows; deployment QA targets what the run receipt authorizes.
Standalone plans: read yoke qa plan run --help. After a lost response, rerun the same command with --resume-execution-id ID per check-in recovery; an unknown reply is not a failure.
Failed CI jobs: follow failed CI job recovery. Never rerun an older run through another tool.

Visual evidence

Run browser checks at their specified sizes. Verify UI fixes with rendered screenshots against the supported deployment context.

Render captures against a universe you name: YOKE_ENV=<local-env> yoke ui up. Seed through that env. Keep admission tokens out of output; add a delay step before a screenshot.
Formal Browser QA runs against its declared post-deploy target. A local preview is a self-check only.

Verification failure ownership

Current-item verification failures belong to the current item. A future item or planned path claim is not a waiver. Fix it here unless a live conflict or explicit operator waiver applies.
Use dependency and claim reconciliation before override. Do not use path-claim-override for a planned future claim; override is last resort and requires a live steering seat covering the project. Route it through yoke say --steering and read yoke claims path override --help.
Update a test this item's change made stale; never weaken a behavior assertion to pass. Route a failure the change cannot explain to steering (yoke say --steering) or the responsible item.
Record the failure, the claim action and the rerun evidence; never report green by pointing at future work.

Health checks

A check declares project_scope, source-tree use, runtimes and capabilities; never branch on a project slug. N/A is not a pass. Project checks live in .yoke/doctor/; see doctor.