Preserve structured failure diagnostics in canonical results.jsonl

[yangziao56]

Problem

BenchFlow persists typed rollout diagnostics in result.json, but the canonical trainer-facing results.jsonl currently drops them.

For example, an idle-timeout rollout can contain this in result.json:

{
  "error_category": "idle_timeout",
  "idle_timeout_info": {
    "reason": "idle_timeout",
    "idle_timeout_sec": 120,
    "idle_duration_sec": 121
  }
}

The corresponding results.jsonl row only exposes the generic error.error = "agent_error" and/or stop_condition = "partial_trajectory". Downstream evaluation and training-data QA therefore cannot distinguish idle timeout, wall-clock timeout, sandbox startup, transport closure, verifier timeout, provider API failure, and suspected API failure without reopening per-rollout result.json files.

The loss happens because _build_rollout_result() already computes error_category, verifier_error_category, and owns a RolloutDiagnostics collector, but its call to _write_results_jsonl() passes only the human-readable error and verifier_error strings.

Proposed bounded scope

1. Add an optional, versioned structured-diagnostics block to results.jsonl under info, preserving:
   • error_category
   • verifier_error_category
   • only present typed diagnostic events from DIAGNOSTIC_REGISTRY
2. Pass the existing categories and diagnostic collector through _build_rollout_result(); do not reclassify human-readable strings in the exporter.
3. Preserve the current error, stop_condition, training_ready, and Prime-SFT fields unchanged.
4. Reuse the existing JSON redaction/non-finite sanitization path so diagnostic text cannot bypass artifact redaction.
5. Preserve backward compatibility: callers that omit the new optional arguments produce the same row shape as today.

Acceptance tests

• Each registered diagnostic class round-trips into the per-rollout row with the correct channel/category.
• Agent and verifier categories remain separate and precedence matches RolloutDiagnostics.category_for_channel().
• Absent diagnostics do not add a block, preserving existing consumers.
• Secrets in raw diagnostic fields are redacted in the written JSONL.
• Non-finite values remain valid JSON.
• Job-level aggregation preserves the block.
• Existing training-readiness and Prime-SFT tests remain green.

Non-overlap

• Diagnostic dataclasses use bare-str fields where Literal/enum belongs; reason constant duplicated across modules #531 concerns enum/literal typing inside diagnostic dataclasses.
• PR fix(diagnostics): unify diagnostic serialization #895 changes diagnostic serialization internals.
• PR Flag chat-only completions (zero tool calls with model output) as a diagnostic #1025 adds a new diagnostic kind.
• PRs fix(result): normalize result artifact timestamps #928 and feat(viewer): reviewer-grade trajectory pages, multi-run browsing, and hf:// dataset sources #1034 address timestamps and metrics.

This proposal only closes the exporter parity gap for already-produced structured diagnostics.

Before opening an implementation PR, could a maintainer confirm that this scope is useful and unowned, the preferred schema location/name, the acceptance path, and whether/at what value this work is eligible for Ziao Yang's FrontierPhysics infrastructure contribution credit?