[AlexSchmidt1999]

What this PR does

GPT-5 and o-series requests start with parameters those models reject. Both OpenAI reply paths now send max_completion_tokens directly. Classic GPT sampling and Gemini's existing temperature stay unchanged. Custom servers keep the legacy-first retry.

The model picker adds GPT-5.4/5.4-mini and current Claude choices, removes retired Claude 3 entries, and excludes families this Chat Completions client cannot use. Anthropic and Gemini model refreshes handle pagination. Gemini reuses the existing image-capable streaming path. Anthropic joins visible text blocks, and provider errors inside successful SSE responses reach the caller.

This updates the Apple clients. Android is unchanged. The parameter change follows OpenAI's Chat Completions reference. The removed Claude models are listed as retired in Anthropic's lifecycle table.

Type of change

• Bug fix

How it was tested

URLProtocol fixtures cover request bodies, model parsing, pagination, retries, and streaming errors. No paid provider requests or health data were sent.

At 6be600c79c4f, the local macOS build and focused StrandTests passed. The iOS Release build, source hygiene, and diff-scoped Apple translation checks passed too. These are local results, separate from GitHub checks.

Checklist

• Generated projects, build output, and credentials are excluded.
• No new dependencies or destructive strap commands.

GitHub validation was rechecked on October 11 at 6be600c79c4f. All 5 registered checks passed, including the macOS app build and tests and the iOS app build.

[AlexSchmidt1999]

I am using NOOP 12.1.0 Release build 456 on my iPhone 17 Pro Max, and the app runs very smoothly in everyday use.

This PR is included in that combined build, together with my other Apple changes and the performance fixes from #2724 by @UtkuDenizAltiok and #2705 by @kavemang. The installed source is d67cceb5.

Both Apple builds and the local app, analytics, database and Python test suites passed for this combined build.