feat(data): task_data schemas for all 108 resources servers

[adil-a]

Problem

PR #2795 ships the task-data schema machinery with only the six example servers covered. The other 108 resources servers still have no machine-checkable definition of their dataset rows.

Change

One task_data.py per resources server. All 108. Each schema was generated from a field-level survey of the server: its committed data shapes, what verify() actually consumes, and its existing wire request models. Each batch was then adversarially reviewed against app.py and the data. Confirmed findings were fixed before this PR.

The conventions, applied uniformly:

• Required-ness mirrors the wire. A field the server's request model requires stays required, even when verify() never reads it. A wire-Optional field stays Optional. Existing datasets keep validating and 422 behavior is unchanged.
• Servers with several row shapes get one model per shape. The field that names the shape picks the model at validation time: ruler2 keys on eval_type, format_verification on verifier.type, asr_with_pc on task_type. Always a field the server already reads, never one invented for the schema.
• Open payloads stay open. reasoning_gym metadata, gymnasium extra fields, per-domain injection environments, and dynamic reference columns are typed as envelopes: the stable keys are declared, the rest stays dict. No false precision.
• Families share models. math_with_judge's heirs, mcqa's heirs, tavily_search's heir, the gymnasium family, and the SWE-instance core import the parent's task_data instead of redefining it.
• Annotations for future tooling. Every field carries a consumed_by tag (verify, metrics, prompt, or provenance). The 14 servers that still use verifier_metadata carry legacy_location markers.
• Pointer and stateless servers (aviary, toolsandbox, code_fim, evalplus, and friends) ship minimal schemas. Their docstrings name where the real task payload lives outside the row.

How this was tested

• Every schema was validated against every committed dataset row that routes to its server, repo-wide. That includes benchmark and environment data living outside the server's own directory. 153 files, 1003 rows, 0 errors, 0 ambiguous duplicate keys. This sweep was run independently after generation. The generators' own reports were not trusted.
• Six servers have no config-declared committed data (their configs point at generated files). Their local tracked example data was validated directly. All clean.
• The CI enforcement sweeps from feat(data): per-server task_data schemas, gym env schema, collate validation #2795 now cover 114 schema files: the dependency-light import check and the extra="ignore" ban. 256 tests pass.
• gym env schema smoke-tested across plain models, unions, and envelope schemas.
• One test fix rides along: the enforcement test now unwraps Annotated[Union[...]] aliases when collecting models.

Stack

PR 2 of the task-data-schema series. Base: #2795 (the machinery). Next: per-server flips of task_data_validation to error, and docs teaching the schema convention.

🤖 Generated with Claude Code

[adil-a]

/ok to test 86d1888

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[adil-a]

/ok to test c6cb251

[adil-a]

/ok to test 191e19c

[adil-a]

/ok to test 45c72e4

[adil-a]

/ok to test b12f47e

[adil-a]

/ok to test 4f83404

[adil-a]

/ok to test 3a75506

[adil-a]

/ok to test 1188068