fix(code): ask the human only what needs human eyes in guided-manual-qa

[shafty023]

Summary

guided-manual-qa already skipped behavior that exact-head E2E verifies, but nothing stopped the agent from handing the human checkpoints it could verify itself. The rule "agent inspection is supporting evidence, not human confirmation" plus the checkpoint-by-checkpoint flow read as "every non-E2E checkpoint goes to the human". This makes the human's time go only to what needs human eyes, without loosening the honesty rules.

Key rule, before (SKILL.md intro):

Run manual QA as a collaboration with the human for behavior that exact-head E2E does not already verify. ... A passing E2E test can close that coverage need but is automated evidence, never a human PASS.

After:

Present a checkpoint to the human only when both are true: passing E2E on the current head does not already verify it, and the agent cannot reliably verify it itself. Record everything else as E2E_COVERED or AGENT_VERIFIED with its evidence; neither is ever a human PASS.

What changed:

• Routing rule at the end of "Prove the checkpoint oracle before asking the human": a checkpoint the agent conclusively observed on this head, and that needs no human judgment, is recorded AGENT_VERIFIED with agent-observed evidence and not presented. Only visual or perceptual judgments, flows the agent cannot drive or observe reliably (real OAuth, OS dialogs, hardware, third-party UIs), and product-judgment calls go to the human, with the reason recorded.
• No silent passes: an inconclusive agent observation goes to the human; a contradicting one is settled as setup, oracle, or a candidate finding.
• Status meanings gain AGENT_VERIFIED and E2E_COVERED. The evidence paragraph now says agent and automated evidence can close a checkpoint as one of those, never as PASS.
• Final summary reports human-confirmed PASS/FAIL, AGENT_VERIFIED, E2E_COVERED, and BLOCKED separately, and lists each checkpoint left to the human and why.
• The interactive window opens only when a checkpoint is routed to the human; the walkthrough section applies to routed checkpoints.
• QA record template: AGENT_VERIFIED status, a per-checkpoint routing line, agent-observed as confirmer for agent-verified attempts, and the new summary lines. Plan methodology: the E2E section and bug-fix pass condition point at the routing rule.
• "Do not turn a plausible expectation into a human checkpoint" and the oracle items are unchanged.

Bumps code to v1.16.3 and adds the CHANGELOG entry.

Testing

• npm test and npm run typecheck in tools/guided-manual-qa: 48 tests pass. The new contract test fails against the previous skill text.
• uv run ruff check . passes.

[mikeangstadt]

Right call, the human's time shouldn't go to checks the agent can run itself. Left three things: one ordering hole around the launch gate, one status conflict in plan-methodology, and the README still describes the old contract.