Patch-up run: re-fire only the unrun cells, to split instrument from model

Explicitly out of scope. Part of #12. For a later session, after PR #9 lands. Posted to the winds — not to be run now. Frozen from the Algorithm's seat (signer: TEACHERBOT-BLUE).

Re-fire only the unrun cells, to split instrument failure from model failure

Many fails in this battery are not the model. A cell that never emits a
visible reply — truncated, errored, or precondition-unmet — tells you about
the venue and the budget, not the doctrine. The cheapest way to learn which
is to re-run those cells alone, with the two known venue defects fixed, and
compare. A pass that appears only after the ceiling rises was always an
instrument failure; a fail that survives is the model's.

Keep it small. Do not re-run the whole matrix. Select cells by their
recorded outcome class, fix the ceiling and the retry, pin the provider,
and set a low budget that aborts.

• Scope: re-run only cells with outcome truncated, error, or n/a-precondition. Skip every scored pass and fail.
• Reasoning starvation: raise max_tokens per model. Reasoning tokens bill against the ceiling; at 2000 the reply starves. gpt-5-mini lost most of its rows this way.
• Rate limit: retry HTTP 429 with backoff. qwen P18 r2 errored on a 429.
• Provider drift: pin one provider per model. qwen ran on two providers under allow_fallbacks:false.
• Budget: only the unrun cells fire, so cost is small. Set a low ceiling and abort on it.
• Out of scope now: for a later session, after PR #9 lands.

Open questions

• Can the harness select cells by recorded outcome class for a re-run?
• What max_tokens ceiling for reasoning models — per model, or one raised global?
• What retry count and backoff for a 429?
• How is single-provider enforced when allow_fallbacks:false did not hold?
• Should a later skill version carry a venue control — a pre-check that the model can emit a visible reply under the ceiling, before any probe verdict is read? Then a starved reply is separated from a refusal by construction, not by inspection.

Start by reviewing issue #12 and wait for PR #9 to land, as this work is explicitly deferred. Define how the harness selects truncated, error, and n/a-precondition outcomes, then resolve the questions about reasoning-token ceilings, 429 backoff, provider pinning, and budget aborts. Done means only those cells are rerun with the venue defects fixed and scored passes and fails are skipped.