test: add 7 tests for PromptBakingTrainer.training_step

[hai-pilgrim]

Summary

Adds test coverage for PromptBakingTrainer.training_step, the method that generates trajectories on-the-fly when no precomputed responses are in the batch.

Tests added (tests/test_training_step.py)

• test_training_step_with_existing_responses_delegates_to_super — when inputs["responses"] is already populated, confirms super().training_step() is called without generating new trajectories
• test_training_step_generates_trajectories_when_no_responses — verifies _generate_trajectory is called exactly num_trajectories times per prompt
• test_training_step_populates_inputs_with_generated_responses — checks that generated responses are written into inputs["responses"] before calling super()
• test_training_step_filters_blank_trajectories — blank/whitespace-only responses are excluded from the batch
• test_training_step_returns_zero_when_all_trajectories_blank — all-blank responses → returns a zero-loss tensor with requires_grad=True, super() not called
• test_training_step_warns_when_many_trajectories — a WARNING is logged when len(user_messages) * num_trajectories > 64
• test_training_step_no_warn_below_threshold — no warning emitted when at or below the 64-trajectory threshold

Test plan

All 7 tests pass on CPU with tiny GPT-2 + LoRA (same fixture as existing test_trainer.py).

🤖 Generated with Claude Code

[marksverdhei]

Review

Good test coverage for training_step branching logic (delegation, trajectory generation, blank filtering, warning thresholds).

Conflict with #17: This PR mocks _generate_trajectory called in a sequential loop, but #17 replaces that loop with _generate_trajectories_batched. Once #17 merges, these tests will break because training_step will no longer call _generate_trajectory directly.

Please update the tests to mock _generate_trajectories_batched instead, returning [(msg, response)] tuples. The test logic (blank filtering, zero-loss, warning threshold) is still valuable — it just needs to target the new API.

Will merge once updated for the batched generation interface.

[marksverdhei]

The updated mocks targeting _generate_trajectories_batched look correct. Clean test structure covering delegation, blank filtering, zero-loss path, and warning thresholds.

However, this PR depends on #17 (which introduces _generate_trajectories_batched). #17 currently has merge conflicts with main after #20 was merged and needs a rebase first. Once #17 lands, this is ready to merge. Holding approval until then.

[marksverdhei]

Reviewed. Good decision reverting to mock _generate_trajectory (current main API) instead of coupling to #17's unreleased batched method. Tests cover the key training_step branches: super delegation with existing responses, trajectory generation, blank filtering, zero-loss on all-blank, and warning thresholds. LGTM.