[0z5a]

When --apply-chat-template is enabled, the standard dataset has already rendered each prompt to a string. Native rollout applied the template again, causing the official Ouro-Thinking tokenizer to fail before the engine call. Apply the template only to structured message lists; encode rendered strings directly.

Incremental draft on #15. The production diff is one condition.

Comparison	Parent #15	Candidate	Speedup
Rendered chat prompt	Fails before engine call	Exact prepared token IDs preserved	Undefined; parent cannot execute
Full online RL	Pending	Pending	Not measured

Validation: 631 CPU checks pass, 6 CUDA skips, 3 known factory deselections. Six new prompt-contract cases cover rendered strings, structured messages and plain text, with thinking on/off; the unchanged parent reproduces both rendered-string failures. An official-tokenizer audit checks all 500 MATH-500 prompts in both thinking modes and two completions per prompt: 2,000 exact input-token matches and known-answer rewards.

The official audit uses an explicit Ray transport seam and canonical answer replay, not model generation or quality measurement. Original production and initial fixture failures are retained. See docs/native_chat_prompt_results.md and benchmarks/results/native-chat-prompt/ for complete scope, source hashes and raw results.