fix(providers): backport ChatGPT Web reasoning transport limits

[diegosouzapw]

ChatGPT Web reasoning preflights currently allow roughly one million inline characters. This backports the conservative 500,000-character ceiling adopted in upstream v6.1.7 (f9ad4ae83a579287105ad822dd0c3e0029b04ef6) for Medium/High and Pro Extra High. A request at the ceiling remains accepted; the next character is rejected before browser dispatch with the existing non-retryable context_length_exceeded error. Multipart staging and final messages use the same policy.

Instant, Pro mode, Luna/Think and token budgets retain their existing limits. The vendor pin remains v4.0.7; the comment records this hunk's upstream provenance. This is a partial behavioral backport for #14194, whose complete runtime refresh and required live ChatGPT smoke remain pending. The 500,000-character ceiling is an upstream safety margin, not a new measurement of the remote service's physical limit.

Validation:

• Repro: the same 17-case fixture records 8 PASS / 9 FAIL on base 7510677f and 17 PASS after the two-constant change. Failures were missing expected preflight rejections, with all existing-behavior controls passing.
• Tests added: tests/unit/chatgpt-web-codex-transport-limits-14194.test.ts invokes the real route resolution and single/multipart preflight functions, covering account/effort combinations, both sides of each boundary, independent token limits and unknown models.
• Focused/related selection: 78/78 across eight files, including the 17 new cases. A no-op loader passes; restoring only the previous constants reproduces the same nine failures. The production files remain unchanged during mutation testing.
• Synthetic HOME, data, vendor state and XDG directories were set before imports; test receipts record no external network attempts, sensitive reads or forbidden child processes. An initial related run failed three Unix-socket setup cases because its temporary path was too long; the same eight files passed after shortening only the runner's temporary directory.
• Formatting, focused lint, test discovery, changelog and the strict repository file-size baseline pass. Core/dashboard typechecks and both complexity gates pass on the same three reviewed file hashes.
• Independent SPEC/Standards review found no blocker. The production delta is exactly two constants and provenance; vendor version/manifest and the other model source are unchanged.

The existing worker/login work in #15935 and #15825 is preserved. This PR does not perform the complete v6.1.7 port or certify a live ChatGPT session. Hosted release CI and the broader #15306 tracking remain separate from these local results.

Refs #14194.

[github-actions]

CI Coverage Report

• Coverage job: skipped
• PR test policy: success

Coverage artifact was not available for this run.