fix(evals): Re-resolve {{uniq}} per retry attempt

[jirispilka]

Problem

runAgentWithRetry in evals/runner/experiment.ts replays a trial's whole prompt after a transient agent error (isTransientAgentError). The {{uniq}} marker introduced in #1423 is resolved once per trial in expandIterations, so the retry asks for the same schedule name. A create case that succeeded on the platform before the transient failure collides with itself on the retry.

#1423's guarantee covers concurrent runs and distinct trials, not a retry of the same trial. Confirmed during that PR's review as a pre-existing gap; the evals/README.md caveat about reading a failure describes it.

Proposed fix

Give each attempt its own suffix, e.g. -t<trial>r<attempt>, by resolving the marker inside the retry loop (or re-substituting from the unresolved item on retry) so a replayed prompt names a fresh resource. isNameFromRun's -<runId>-t match keeps working because the run id and -t delimiter are unchanged. Add a unit test that two attempts of one trial resolve to different names.

Context

• Mechanism this extends: fix(evals): Make schedule eval runs safe to run in parallel #1423 (evals/run_id.ts, substituteUniqMarker).
• Live schedule eval run showing the current per-trial naming: https://langfuse.apify.dev/project/cmshkde21000krg07shb46d8g/datasets/cmtt5bb54020itx07em540k7g/runs/9927222dca474925
• Parent issue: Schedule evals share mutable account state, so runs can't be parallel #1394.

Drafted by Claude Code during the #1394 shepherd run.