[evals] Add GH Archive structured-output PPL evals

[dlwh]

Normalize a small held-out GH Archive slice into stable masked JSON text records split by event type, and wire an opt-in raw validation registry plus perplexity-gap experiment entrypoint. Add synthetic tests for filtering, masking, serialization stability, optional WorkflowRun handling, and step hash coverage.

Fixes #5098

[claude]

Claude finished @dlwh's task in 7m 4s —— View job

---

Code review

🤖 No issues found. Checked for bugs and CLAUDE.md compliance.

Reviewed:

• experiments/evals/gh_archive_structured_output.py
• experiments/exp_model_perplexity_gap_gh_archive_structured_output.py
• lib/marin/src/marin/datakit/download/gh_archive.py
• tests/evals/test_gh_archive_structured_output.py

Against: root CLAUDE.md / AGENTS.md, experiments/AGENTS.md, lib/marin/AGENTS.md.

Candidate concerns that were investigated and dismissed:

• make_gh_archive_step thin wrapper — matches the established pattern make_uncheatable_eval_step in the same directory; not an unjustified abstraction.

• Dual-optional raw_root / gh_archive_raw params — the CLAUDE.md X | str rule targets a single union parameter, not two mutually-exclusive optionals.

• Hash-attrs test — not tautological; it enforces that window/cap fields participate in the step's identity, so dropping any one would produce a real regression.

• Masking determinism, idempotency, and cap enforcement — sort_keys=True + sorted parse_qsl/urlencode + ExitStack-ordered atomic renames are internally consistent.

• Check if PR needs review

• Identify relevant CLAUDE.md/AGENTS.md files

• Summarize PR changes

• Run parallel review agents

• Validate flagged issues

• Post review findings
  • Branch: agent/20260423-fix-5098