test(campaign): follow retained attempt receipt paths

[drewstone]

Change

Update the eleven existing controls that still read retired flat failure receipts.
They use Eval's existing campaign attempt-directory function and the recorded latest-attempt identity.
Error classification, spend attribution, sibling draining, retry limits, cancellation and holdout refusal checks remain.
Remove the old “writes spans.jsonl per cell” case, which only counted cell directories and never read a trace.

Evidence

• Actual final fix(campaign): preserve artifacts and traces for every attempt #908 CI failure: run 36898653521. All eleven failures used retired flat receipt paths.
• Actual installed public Eval 0.203.1 reproduces original artifact and trace loss under retry and resume. Retained at /mnt/traces/eval-release-02032-20261001/public-02031-baseline.
• Frozen install, Biome checks and normal commit-hook typecheck qualify this narrow edit. No new tests or local suites.
• Production source, package versions, dependency locks and workflow commands are unchanged.
• Exact public 0.203.2 campaign qualification follows the immutable publisher; this PR does not claim it already ran.

Root assigned these two existing test files to this repair lane after #908 merged. The prepared release remains owned by existing PR909.

[tangletools]

✅ Auto-approved PR — 0de4005b

Blanket team auto-approval is intentional. The merge gates are CI and codex-p1.
No automated review runs on this PR. This approval rests on the rule above alone.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-10-01T17:59:45Z}

[drewstone]

Independent review: two existing test files follow Eval’s maintained cellAttemptDirectory and persisted latest-attempt identity. Error classification, cost contents, sibling draining, native retry bounds, cancellation and held-out refusal controls remain. The deleted spans case only counted directories and did not check a trace. Actual public203.1 overwrite and source908 consumer retention are retained; no production source, dependency or workflow change. Fresh actual-target merge is clean. This is assertion maintenance, not the public203.2 execution proof.