fix: remove orphaned outputs reference, clean up test suite (#206)

Bug: `gen_len = outputs[0].shape[0] - ...` was outside the if/else
block — `outputs` only exists in the unconstrained path, crashing
the constrained path with UnboundLocalError.

Test cleanup: removed source-code-grepping tests (brittle, test
implementation not behavior). Kept 31 tests that verify behavior:
- Regex matching (valid/invalid inputs)
- DFA quantifier limits
- Outlines API contract (imports, dispatch types, image wrapper)
- Generator cache sentinel logic
- Task rotation (loading, explicit IDs, rotation coverage)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: abrichr

openadapt_evals/training/standalone/trainer.py

@@ -252,7 +252,6 @@ def _collect_rollout(self, task_id: str, instruction: str) -> Rollout:
                     skip_special_tokens=True,
                 )
                 gen_len = outputs[0].shape[0] - inputs["input_ids"].shape[1]
-            gen_len = outputs[0].shape[0] - inputs["input_ids"].shape[1]
             action = parse_vlm_output_to_action(decoded, screen_size=self._config.screen_size)
 
             if gen_len >= self._config.max_new_tokens - 1: