fix(datasets): validate, gate, and drain cached file fetches; single retry owner for banking

[devin-ai-integration]

TL;DR

Hardens fetchCachedTextFile (added in #75/#78) against four issues found by Devin Review on the downstream subtree sync (openrouter-web#41251): cache poisoning by malformed 200s, ignored BENCH_DATASET_CACHE_DISABLE, unreleased error bodies, and multiplied retries in the banking dataset.

What changed?

• cached-file.ts
  • New optional validate?: (text) => string | undefined on CachedTextFileRequest; a rejected body fails with CachedFileError (no status, so not retried) and is not written to the cache.
  • New jsonTextValidator("object" | "array") helper.
  • store.enabled is now honored before both readJson and writeJson (mirrors the HF page cache).
  • Non-2xx responses drain their body (ignore(response.text)) before failing.
• tau-bench-airline/environment.ts: db.json fetch uses jsonTextValidator("object").
• tau3-bench-banking/environment.ts: ensureBankingData/ensureBankingTasks accept an optional RetryConfig and validate db.json (object) / tasks.json (array).
• tau3-bench-banking/dataset.ts: drops the outer retry(hfFetchRetrySchedule(retryConfig)) and threads retryConfig into ensureBankingTasks, so maxRetries bounds origin requests exactly once.

Why?

The tau file cache keys on immutable URLs with no TTL, so a single malformed 200 (e.g. an HTML rate-limit page) would be served forever from disk/GCS. Banking retries were (inner 3 default) × (outer maxRetries), so maxRetries=0 still made 4 requests.

How to test

bun test src/datasets/cached-file.test.ts src/benchmarks/tau3-bench-banking

New cases: validator rejects a 200 body without caching or retrying; disabled store sees zero reads/writes; failed response bodyUsed === true; banking maxRetries: 0 makes exactly one request.

Benchmark impact

No score change. Failure mode changes: a malformed dataset download now fails the run instead of silently caching bad data.

Reviewer focus

• ensureBankingData/ensureBankingTasks signature gains an optional trailing param — backward compatible; the solver call sites still use defaults.
• Whether draining (vs. cancelling) the error body is acceptable for the FetchHttpClient.

Checklist

• Tests cover changed behavior
• Public API or configuration changes are backward compatible, or the break is documented
• Benchmark changes document dataset provenance and licensing
• No credentials, private results, or restricted dataset contents are included
• Documentation is updated where needed

Link to Devin session: https://openrouter.devinenterprise.com/sessions/949ffdf48c7843febc66e0c2d8dc8486
Open in Devin Desktop: https://openrouter.devinenterprise.com/desktop/session/949ffdf48c7843febc66e0c2d8dc8486?variant=devin
Requested by: @abhinav-pola

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR that start with 'DevinAI' or '@devin'.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

---

Original prompt from Abhinav

SYSTEM:
<latest_message>
Abhinav Pola (U090K0G7JF3) [ts=1788974411.868429]: @Devinsync benchmark harness subtree
</latest_message>

=== BEGIN THREAD HISTORY (in #ecosystem-agents) ===
Abhinav Pola (U090K0G7JF3) [ts=1788974411.868429]: @Devinsync benchmark harness subtree
=== END THREAD HISTORY ===
Channel ID: C0BU53A7VEH
Thread URL: https://openrouter.slack.com/archives/C0BU53A7VEH/p1788974411868429?thread_ts=1788974411.868429&amp;cid=C0BU53A7VEH

The <latest_message> is the message that you should use to guide your goals + task for this session, and you should use the rest of the slack thread as context.
A [ts=...] marker on a Slack message is that message's timestamp. To act on a specific message with the slack tool (e.g. adding an emoji reaction via the reaction command), pass that value as timestamp along with the Channel ID — no extra lookup call is needed.