fix: use max_completion_tokens for gpt-4.1+, gpt-5.x, and o-series models

[YuHuang0525]

Problem

Newer OpenAI models — gpt-4.1, gpt-5.x, o1, o3, o4 — reject the legacy max_tokens parameter with HTTP 400 and require max_completion_tokens instead. This means any user who sets LLM_MODEL=gpt-5.2 (or any other newer model) gets an immediate 400 error on every LLM call, breaking scoring, tailoring, and cover letter generation entirely.

Relevant code before this fix (llm.py _chat_compat()):

payload = {
    "model": self.model,
    "messages": messages,
    "temperature": temperature,
    "max_tokens": max_tokens,   # ← rejected by gpt-4.1+, gpt-5.x, o-series
}

Fix

Detect the model prefix at call time and send the correct parameter:

_new_param_models = ("gpt-4.1", "gpt-5", "o1", "o3", "o4")
if any(self.model.startswith(p) for p in _new_param_models):
    token_param = {"max_completion_tokens": max_tokens}
else:
    token_param = {"max_tokens": max_tokens}

• All other providers (Gemini compat, Gemini native, local/Ollama) are unaffected — they continue using max_tokens as before.
• The native Gemini path already uses maxOutputTokens and is untouched.
• No behaviour change for existing gpt-4o, gpt-4o-mini, or local model users.

Testing

Verified manually with gpt-5.2 — the 400 error is resolved and completions return successfully after this change. No existing automated tests cover _chat_compat() directly.