Send max_completion_tokens to OpenAI reasoning models

[TheGreatAxios]

Summary

• First-party OpenAI reasoning models (gpt-6-astra, gpt-5.4, gpt-5.4-mini, o3, o4-mini) now send max_completion_tokens instead of max_tokens, which their endpoint rejects
• Requirement declared per model on the OpenAI API-key preset catalog entry, never inferred from name prefixes; the quirk attaches per-source and follows the first-party endpoint, so relays serving the same model names keep max_tokens

Verification

• bun test packages/first-class-providers/src/providers.test.ts src/config/inference-sources.test.ts: 23 pass, 0 fail
• bun run check (lint, typecheck, build, guarded suite): exit 0; 7079 pass, 0 fail across 485 files
• Bug-first proof: with the attach stashed, reasoning models emit no max_completion_tokens; the fix restores it

Manual release-gate (live smoke, CI stays request-body level)

• One-turn live check against the OpenAI API-key preset with a reasoning model: confirm the turn succeeds and the request carries max_completion_tokens. Needs a real API key, so it stays a manual step — never commit secrets or network-dependent tests.

Fixes CL-7785

[linear-code]

CL-7785