feat(provider): OpenAI Responses background mode (sa-73)

[topcheer]

Summary

Implements OpenAI Responses background mode (official guide: platform.openai.com/docs/guides/background, openbeta) for the stateless-replay Responses adapter, so multi-minute reasoning turns (GPT-5.x Pro, codex-max) no longer depend on a single held-open HTTP connection.

Problem

ggcode's Responses client was purely synchronous: POST /responses blocked until the full response finished. GPT-5.x Pro / codex-max runs of 5+ minutes repeatedly died on idle timeouts and dropped connections (recurring pattern in sa-66/sa-71).

Implementation

• Create with background: true -> API returns a Response object with status queued immediately and executes server-side.
• Poll GET /responses/{id} with exponential backoff (1s -> 10s cap) until a terminal state (completed/failed/incomplete/cancelled); tolerates up to 5 consecutive transient poll errors.
• Cancel on context cancellation via idempotent POST /responses/{id}/cancel (best-effort, fresh 5s context) so a dropped turn stops burning tokens.
• Streaming: ChatStream in background mode diffs text across poll snapshots for incremental deltas; reasoning items / function calls / hosted web_search items are replayed from the terminal payload - the stateless encrypted-reasoning loop (sa-54) and hosted-tool echo (sa-62) work unchanged.
• Config: opt-in per endpoint via responses_background: true, wired EndpointConfig -> ResolvedEndpoint -> registry (both the openai-responses case and the URL-sniff fallback).
• Chat() non-background behavior is untouched (decode path refactored into shared chatResponseFromPayload).

Test plan

• 9 new httptest-based tests: request flag, poll-to-completion, transient-error retry, failed/cancelled/incomplete terminal states, cancel-on-context-cancel, incremental stream deltas, config wiring
• go test -race -tags goolm clean on background tests
• Full internal/provider + internal/config suites pass (-p 1 -parallel 2, -tags goolm)
• linux/amd64 + darwin/arm64 go vet -tags goolm ./... clean; gofmt clean
• Docs: new "Background mode" section in docs/guide/providers.md

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

[topcheer]

复审通过 ✅（三重点核销）

① 轮询面：指数退避 1s 起倍增至 10s 帽（长任务不空转 API）+连续 5 错熔断（持续端点失败浮出而非永轮——瞬态断连容忍由 background 模式存在意义注释论证）+ctx 取消优先分支（先 best-effort cancelResponse 再返回——dropped turn 不烧 token）。

② 前后台消息序：onPoll 增量快照（每次 poll payload 与上次 diff 出新 text 流式发）——文本顺序=服务器产出顺序（快照单调追加）；后台模式无 tool_call 前后台竞态（terminal 才返回完整 payload——工具调用经终态统一交付）。

③ 成本归属：终态 payload 的 usage 随正常 Chat 返回路径进本次 run 的 TokenUsage——后台 token 计入发起 run（无归属漂移）；cancel 场景 usage 丢弃（服务器侧取消语义）。

opt-in 配置门（responses_background yaml）+非后台路径零变。CI 绿。可合并。

[topcheer]

复审已过（见 review：指数钳 1s→10s+5 错熔断+ctx 取消先 cancelResponse/增量快照流/usage 归属发起 run）。与 main 冲突——#2515/#2517 先合并动了 openai_responses.go 相邻区（responsesRequest struct 的 Reasoning/Text/Background 字段区+post 方法）。分支 sa-73-background 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。