feat(provider): add service_tier processing-tier control for openai endpoints

[topcheer]

Summary

Adds service_tier (sa-81) support on openai and openai-responses protocols — the OpenAI processing-tier parameter that trades cost against latency (flex = cheaper/slower, priority/fast = Fast mode, the 2026-07 rename of Priority processing). Follows the exact plumbing pattern of the existing text_verbosity / reasoning_effort controls.

Changes

• config: new service_tier endpoint field (service_tier: yaml key), trimmed into ResolvedEndpoint alongside the other hints
• provider: new ServiceTierProvider interface; registry wires it for openai, openai-responses, and the /responses URL-sniff path
• Allowlisted setters: unknown values are ignored (existing setting preserved) so a typo cannot silently flip an expensive production tier; valid values auto | default | flex | priority | fast | scale
• openai.go: tier rides Chat Completions (non-stream + stream) requests; transparent one-shot degrade-retry when a relay rejects the parameter (same pattern as retryWithoutReasoningEffort)
• openai_responses.go: tier rides Responses requests with graceful degrade identical to text.verbosity
• fallback.go: tier forwarded across fallback chains via forEach
• CloneWithModel carries the tier across model swaps
• docs: new "Service Tier" section in docs/guide/providers.md

Anthropic/Gemini/Copilot endpoints are unaffected (no interface implemented → field ignored, documented).

Tests

10 new tests, all passing:

• config: yaml resolution + trim + empty default
• openai: setter validation, applyServiceTier, clone-carries-tier, error classification
• openai-responses: setter validation, wire injection (service_tier in request body), omission when unset, degrade-retry against a gateway returning 400 Unknown parameter: 'service_tier'

Verification: gofmt clean, go vet -tags goolm clean, go build -tags goolm ./... OK, full internal/provider + internal/config regression suites green, pre-commit/pre-push hooks passed.

Co-Authored-By: ggcode noreply@ggcode.dev

[topcheer]

复审通过 ✅（四重点核销）

① 显式 6 值 case 表+未知忽略（typo 不翻转昂贵档，三 setter 家族形态一致）。② 与 #2528 无耦合——tier 重试 req 级内层/callPolicy 外层正交；与 effort 重试串联（双 drop 最多 3 次请求上界合理）。③ Fallback 链式广播同 verbosity 形态。④ CloneWithModel 值继承（端点属性语义）。retryWithoutServiceTier Param 精确→消息兜底两态。

CI 绿。可合并。

[topcheer]

复审已过（见 review：①显式 6 值 case 防翻转+未知忽略（typo 无声面=effort/verbosity 家族既有取舍一致）②与 #2528 正交——tier 重试 req 级内层/callPolicy 外层，rebase 时接口注册顺序无耦合（不同 setter 名）③fast 拒收走 retryWithoutServiceTier 统一降级（Param/消息两态判别覆盖老网关）④fallback 广播+Clone 值继承 ⑤scale 透传由 OpenAI 域判定（未知 tier 服务端拒→降级路径兜底））。与 main 冲突——#2529 等先合并动了 openai.go 相邻区。分支 sa81-service-tier 有归属者本地 worktree 在途——请归属者 rebase main 解冲突，解后 CI 绿即合并。不代解。