[osolmaz]

Summary

A review of #37, by a reviewer that did not write it, found bugs and leftovers of the old interface.
This PR fixes them.
The main bug: the viewer showed "The model is answering…" for good on a step whose request failed.

What Changed

Most fixes are small. Two change behavior in the examples.

• Viewer:
  • A step without answers at a failed run or a stopped case is marked failed, with its reason (Frame.failed).
  • A pick is found by the facts that picking writes. The new core helper pickKeys does this, and the replay check uses it too. A choice question that is no pick, such as the README's team, is no longer charted as a pick. The unused FrameChoice.key is removed.
• WikiRace: a dead end (no link or section left) ends the race without a request that offers only "give up". The outcome is unfinished with the reason dead end. Before feat!: make every step a set of questions #37, such a race also ended without a request.
• rubric-judge:
  • When the gathering steps reach the judge's limit, the state now records that the gathering ended. So a fork before the checks, under a judge with a higher limit, asks only the checks.
  • When only stop is left to pick, the checks come at once.
  • The outcome loses steps, which had a second meaning next to result.steps. The README says to fork at steps - 1.
• Docs and leftovers: the trace and result comments in src/cli/layout.ts, the package description, P(path) in docs/MATH.md and the plan (choice answers only, because a bool answer names no label), the viewer README, the README's list of the three examples, and a test that named the removed --state-from flag.

Testing

All tests use the fake decider; no model was called.

• pnpm check at the root passes. Tests: root 71, viewer 39, chess 75, rubric-judge 49, WikiRace 20.
• New tests: pickKeys; the failed-step panel; a thrown case's step; the WikiRace dead end; the judge's fork after its limit; the judge with only stop left.

Risks

The risk is low. These changes follow the design of #37, and the full check passes.

• rubric-judge states written by feat!: make every step a set of questions #37 at the gathering limit have gathered: false. A fork of such a trace under a judge with a higher limit can still gather more. No real runs exist with feat!: make every step a set of questions #37 yet.