Reduce GRPO trainer memory

[bingran-you]

What changed

• Reduces the sampled-logprob model-bridge context cap from 24,576 to 16,384 tokens.
• Leaves baseline, SFT, gate, and final evaluation at the full 49,152-token context.
• Sets PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True at CLI startup unless the operator supplied another allocator configuration.

Why

The 24,576-token cap prevented the trainer's forward/logprob OOM, but a later variable-length batch still OOMed during backward due to fragmentation: PyTorch had 10.41 GiB reserved but unused and attempted an 11.66 GiB allocation.

A 16,384-token GRPO context reserves 4,096 tokens for generation and bounds the sampled prompt to 12,288 tokens. Expandable segments reduce fragmentation across the highly variable data-agent trajectories. Evaluation behavior remains unchanged.

The failed GRPO stage still had zero reward variance, loss=0, and grad_norm=0, so restarting from the saved SFT checkpoint loses no learned 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 12,288 prompt tokens and returned a sidecar with the same exact prompt-ID count.