test(live): treat error as a terminal and skip models this account cannot use

[kaizhou-lab]

run_backend_set_model waited only for a finish frame, so a turn that ended in an error burned the full 300s and then reported "the turn did not finish" — a message naming the wrong problem. On 2026-09-07 that very nearly went into the record as a codex 0.153.4 regression.

What is actually happening

Probing the frames settled it. The turn does terminate, and the reason does reach the user:

frames: ["start", "tips", "tips", "tips", "tips", "tips", "error"]

tips:   "Reconnecting... n/5 — unexpected status 403 Forbidden: user not allowed to
         access model. This user can only access models=['standard'].
         Tried to access gpt-6-astra"

error:  {"code":"USER_LLM_PROVIDER_PERMISSION_DENIED",
         "ownership":"user_llm_provider",
         "detail":"unexpected status 403 Forbidden: user not allowed to access model…"}

There is no product defect. An earlier note suspecting the app-server swallowed the 403 was wrong and is retracted — the backend already classifies this as the user's provider's problem and says so, with the model name and the allowed set.

Two changes, both to the test

1. error is a terminal. The wait ends on either terminal, and the failure message carries the backend's own detail plus the warnings that preceded it. Failure time on this account: 314s → 28s, and the message now names the entitlement rather than a phantom hang.

2. Try every candidate model, not just the catalog's first. The catalog lists models an account may not be entitled to, and its order is not stable: 0.153.4 led with gpt-6-astra (403 here), 0.151.0 led with gpt-5.6-sol, and two runs tonight disagreed with each other. Picking the first made the test's subject vary with the release — so an A/B across versions was not comparing the same thing at all, which is precisely how the false regression arose.

The part that had to be discovered by running it

A turn ending in error leaves the conversation with no active agent, so the next switch 404s:

{"success":false,"error":"No active agent for this conversation","code":"NOT_FOUND"}

The retry therefore re-ensures the runtime first, through the same endpoint the test already uses to start it. This was not assumed — the first iteration attempt failed exactly there, and an earlier attempt at this fix (2026-09-07) was withdrawn unshipped for the same reason before the cause was understood.

Verification

run	result
codex 0.153.4, single test	ok, 26.71s — gpt-6-astra refused and recorded, gpt-5.6-sol completed
set_model_takes_effect, all backends	2/2 ok (claude + codex; agy has no such test)

The helper is shared across backends, so both were run rather than only the one being fixed.

Clippy clean, fmt clean.

Relation to the codex bump

This is one of the two things keeping codex's gate B from being green. The other is #976 (the plan tool became opt-in). Neither depends on the other; both are needed before VERIFIED_CODEX_VERSION can move off 0.151.0.

Test-only change — left for human review, no auto-merge.