refactor(gpqa): drop structured runner; ship configs/gpqa/run.sh

[ishandhanani]

Summary

Ports GPQA to the configs/<bench>/ script-based pattern that AIME moved to in #91. Closes step 1 of follow-up issue #92.

Same nemo-run unquoting hazard motivates the move: any benchmark passing Hydra ++overrides through ns eval is one backslash-bearing extract_regex away from silently-broken evaluation. Script-based path is immune.

What ships

Removed

• src/srtctl/benchmarks/gpqa.py (62 LOC, sglang.test.run_eval-based runner)
• src/srtctl/benchmarks/scripts/gpqa/bench.sh (53 LOC)
• BenchmarkType.GPQA enum
• gpqa registry assertion in tests/test_benchmarks.py
• GPQA row + section in docs/config-reference.md
• gpqa mention in examples/example.yaml and docs/architecture.md tree

Added

• configs/gpqa/run.sh — orchestrates ns prepare_data gpqa + ns eval --benchmarks=gpqa:\$REPEAT against localhost:8000/v1 (the in-job dynamo frontend). Tuning knobs (MAX_TOKENS, REPEAT, NUM_THREADS, TEMPERATURE, TOP_P, SEED, DATASET, MODEL) env-var overridable; defaults match the upstream reasoning-eval reference (max_tokens=400000, repeat=32, temperature=1.0, num_threads=512). Warns at startup if HF_TOKEN is unset (GPQA Diamond is HF-gated).

Updated

• docs/accuracy.md GPQA section → script-based runbook with HF gating note + pointer back to AIME's reasoning-mode env vars
• src/srtctl/benchmarks/__init__.py — drop gpqa import + __all__ entry
• 6 existing type: "gpqa" recipes migrated 1:1 to type: custom + bash /configs/gpqa/run.sh, preserving their MAX_TOKENS / REPEAT / NUM_THREADS as env-var overrides
• 2 h200 recipes' commented-out type: "gpqa" block replaced with a one-line pointer to the new docs

Recipe shape (from new docs section)

benchmark:
  type: custom
  container_image: nemo-skills    # alias from srtslurm.yaml `containers:`
  env:
    OPENAI_API_KEY: "EMPTY"
    HF_TOKEN: "\${HF_TOKEN}"      # REQUIRED: GPQA Diamond is HF-gated
    # MAX_TOKENS / REPEAT / TEMPERATURE / SEED / etc. all overridable
  command: |
    bash /configs/gpqa/run.sh

Backward compatibility

None. Recipes with type: gpqa will fail schema validation. Migration is 1:1 to the block above. The 6 in-tree GPQA recipes are migrated in this PR.

Test plan

• `make check` — 611 passed, 2 skipped (matches pre-change baseline; drops the one gpqa registry assertion)
• (post-merge) End-to-end on a GB200 cluster: validate `ns prepare_data gpqa` succeeds with `HF_TOKEN` plumbed through `benchmark.env`, eval lands real `metrics.json`

Follow-up

MMLU / longbenchv2 / gsm8k still to come per #92 — separate focused PRs in the order the issue lays out.

🤖 Generated with Claude Code

[qiching]

LGTM