[mvanhorn]

摘要

Fixes #1165

After speech recognition, polish on an OpenAI channel still runs at the normal service tier, so the rewritten text waits a few seconds before it is inserted. The low-latency choice already offered for Ark is unavailable on OpenAI, so there is no way to pay a higher rate for a shorter wait.

修复 / 新增 / 改进

OpenAI channels gain a per-channel Fast mode switch.

• With it on, Chat Completions and Responses bodies include a top-level service_tier of fast beside the existing model and reasoning fields.
• An off switch, the Messages format, and custom, opencode, and openai-compatible providers leave the field out. Ark still sends fast on every format.
• Desktop and Linux settings use OpenAI-specific copy that states the higher price and links to the official pricing page.
• Blocking and streaming diagnostics record the tier sent and the service_tier on the response, or absent when none is returned.
• Leaving the OpenAI provider clears the saved value, and an unsupported tier still surfaces as the provider's invalid response.

兼容

• 不包含：service_tier for the Messages format and for custom, opencode, and openai-compatible providers.
• 对现有用户 / 本地环境 / 构建流程的影响：service_tier is sent only when Fast mode is switched on for an OpenAI channel. Ark behavior is unchanged.

测试计划

Tests for the openless LlmProtocolConfig change live in openless-all/app/src/pages/settings/LlmProtocolFields.test.ts.

• 命令：
• 结果：
• 证据路径：

[mvanhorn]

Merged beta into the branch as 4bab17a. The only conflict was in linux-egui/src/i18n.rs, where the two Fast mode catalog entries sat next to the reindented messagesThinkingLabel entry; I kept both and matched the new indentation. cargo test -p openless-core --lib passes (1421 tests). The linux-egui crate depends on Wayland and cpal, so I could not build it on macOS, and CI covers that part.