OpenAI 400 Unsupported parameter max_tokens: use max_completion_tokens instead

# OpenAI 400 Unsupported parameter max_tokens: use max_completion_tokens instead

## The symptom

`400 Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.` Fires on o-series and GPT-5 reasoning models. Code that worked on GPT-4o breaks the moment it points at a reasoning model.

## Confirm the cause

`max_completion_tokens` replaced `max_tokens` for newer models, and the rename is semantic: reasoning models generate internal reasoning tokens that are billed but never returned, so "tokens generated" and "tokens you see back" are no longer the same number. The new parameter works on all current models, making it the portable choice.

The trap: the budget covers ALL output tokens including hidden reasoning. A small budget can be eaten entirely by internal thinking, leaving visible content empty with `finish_reason: "length"`. An empty answer with finish_reason length on a reasoning model usually means the budget starved, not that the model had nothing to say.

## The fix

1. Send `max_completion_tokens` everywhere, never `max_tokens`. One code path covers GPT-4o, GPT-3.5-era, and reasoning models.
2. Size it generously for reasoning models. Empty content plus finish_reason length means raise the budget substantially or lower `reasoning_effort` to cap thinking overhead.
3. Grep every call site when migrating, including wrappers and SDK defaults. The ones you miss 400 at 2am.

## Verify the fix

Call the reasoning model with the new parameter and confirm 200. Deliberately test the edge: a tiny budget should produce the empty-content plus finish_reason length signature, and your code should read that as "budget too small", not "model broken".