Cap GRPO sampled-logprob context

[bingran-you]

What changed

• Adds --max-logprob-context-tokens, defaulting to 24,576, to the OpenCode model bridge.
• Applies the stricter cap only when the request explicitly asks for sampled-token logprobs.
• Keeps the ordinary 49,152-token context for baseline, SFT, gate, and final evaluation.
• Reuses the existing tool-output-only context fitting, so system/user instructions remain intact.

Why

After raising the aggregate rollout limit, a valid ~37k-token data-agent trajectory reached TRL but OOMed during trainer-side policy-logprob recomputation. PyTorch attempted a 34.33 GiB logits allocation on top of 53.8 GiB already used.

Evaluation can serve the full context on the inference H100, but GRPO must also recompute logits on the trainer H100. A 24,576-token sampled-logprob context bounds that trainer path while preserving the full evaluation context and exact sidecar token IDs.

The failed GRPO run had zero reward variance, loss=0, and grad_norm=0 through six completed steps, so the stage can restart from the saved SFT checkpoint without losing a policy update.

Validation

• 224 package contract tests pass.
• 26 focused bridge/CLI tests pass.
• Ruff check/format, Python compilation, and git diff --check pass.
• Live replay of the known 59k-token follow-up with logprobs=true fitted to exactly 20,480 prompt tokens, returned from the real Qwen3.5-9B TRL server, and produced a sidecar with the same 20,480 exact prompt IDs.
• Ordinary evaluation requests still use the separate 49,152-token cap.