Fix malformed <|im_start> token in Phi4Reasoning prompt style

[JSap0914]

Bug

Phi4Reasoning.apply builds the first ChatML turn with a malformed special token: it emits <|im_start>system (missing the closing |) instead of the well-formed <|im_start|> token that Phi4 and every other ChatML-based prompt style use.

return f"<|im_start>system<|im_sep|>{sys_prompt}<|im_end|><|im_start|>user<|im_sep|>{prompt}<|im_end|><|im_start|>assistant<|im_sep|>"
#        ^^^^^^^^^^^ missing the closing pipe; the other two occurrences are correct

Because <|im_start> is not the registered special token, the Phi-4-reasoning tokenizer does not collapse the system turn into the single <|im_start|> token, corrupting the chat template for that model.

Fix

Use the well-formed <|im_start|> token for the system turn, matching Phi4 and the two other occurrences in the same template.

Verification

Added a regression test in tests/test_prompts.py:

$ pytest tests/test_prompts.py::test_phi4_reasoning_uses_well_formed_im_start_token -v
test_phi4_reasoning_uses_well_formed_im_start_token PASSED

The test asserts the rendered template contains no malformed <|im_start> marker, starts with <|im_start|>system<|im_sep|>, and uses the well-formed token exactly three times.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.