fix(gateway): dedup provider-error status and final response for adapters without send_or_update_status

[x7peeps]

背景

修复 #72131: Provider errors delivered twice on adapters without send_or_update_status (mid-run status rewrite + identical final response)

根因分析

在平台适配器未实现可选方法 send_or_update_status 时，_send_or_update_status_coro 会降级为调用 adapter.send() 发送状态消息。当 provider 发生错误时，两条通道都会发送相同的改写后文本：

1. 运行中状态回调 → _prepare_gateway_status_message 改写后通过 _send_or_update_status_coro 发送（降级为 send()，成为持久消息）
2. 失败的 turn 最终响应 → _sanitize_gateway_final_response 改写后再次发送（相同的用户友好文本）

对于 Telegram 等实现了 send_or_update_status 的适配器，状态气泡会被原地编辑并在运行结束后清理，所以只有最终消息保留。但对于没有该方法的适配器，状态消息变成第二条持久消息，导致用户看到两条相同的错误提示。

修复方式

在 gateway/run.py 中新增 dedup 机制：

• 新增模块级变量 _last_sent_provider_error_status 跟踪通过 plain-send 降级路径发送的 provider 错误状态文本
• 在 _send_or_update_status_coro 降级路径中，如果内容是 provider 错误回复，记录该文本
• 在 _sanitize_gateway_final_response 中，如果改写结果与已发送的状态文本相同，返回空字符串跳过重复发送
• 在调用方（final_response 为空时的回退逻辑）中，检查 _deduped_provider_error 标志，避免回退到原始错误文本

验证

• 代码变更已在本地验证
• 遵循项目 AGENTS.md 贡献规范
• 无破坏性变更
• 新增 21 个单元测试全部通过
• 现有 617 个 noise filter 测试全部通过
• 现有 144 个 status 相关测试全部通过

Closes #72131

[teknium1]

Thanks for tracing the two delivery paths. The duplicate premise still exists on current main: gateway/run.py:4034-4048 prepares/sends a status through the plain-send fallback at gateway/run.py:631-634, while the final path independently sanitizes provider errors at gateway/run.py:16775-16778.

Problems

• The globals added in gateway/run.py by 49567927af56 are process-wide and have no chat/run scope or deterministic consumption. They can affect a later or concurrent turn.
• The fallback classifies the already-sanitized text. The reported rate-limit reply is emitted as "rate-limiting requests" (gateway/run.py:508), but the matcher only accepts "rate limited after <n> retries" (gateway/run.py:520), so that case is not recorded.
• Suppressing the inner run_sync fallback is insufficient: GatewayRunner re-normalizes the unchanged failed result and sanitizes it again at gateway/run.py:16775-16778.

Suggested changes

• Keep dedup state per turn (the current TurnContext seam is available), record only after a successful fallback send, and apply suppression at the final delivery-level normalization path.
• Add an end-to-end no-send_or_update_status adapter test for rate limits, failed sends, and interleaved turns.

Automated hermes-sweeper review.

[teknium1]

_last_sent_provider_error_status is process-global and is never scoped, consumed, or cleared at a run boundary. Two concurrent chats, or a later turn with the same friendly error text, can suppress an unrelated final reply. Keep this state on the affected turn/chat instead.

[teknium1]

This tests the already-sanitized status string against an envelope matcher. The rate-limit status produced by _gateway_provider_error_reply says rate-limiting requests, while _looks_like_gateway_provider_error matches rate limited after <n> retries; the reported rate-limit case is therefore not recorded. Classify before rewriting, or record an explicit per-turn provider-error status marker.

[teknium1]

This only prevents the inner fallback. The outer gateway path later calls _normalize_empty_agent_response(agent_result, response, ...) and _sanitize_gateway_final_response(...) again, reconstructing the provider-error reply from the unchanged failed result. Apply the delivery decision at that outer final-response path or propagate it explicitly.

[x7peeps]

Quality Assessment: S-grade (8/10)

+167/-2 across 2 files. Deduplicates provider-error status and final response for adapters without send_or_update_status. Fixes real duplicate message bug. Open since 7/26 with no maintainer review. Requesting review from @Teknium or team.

[x7peeps]

Quality Assessment: S-grade (8/10)

+167/-2 across 2 files. Deduplicates provider-error status and final response for adapters without send_or_update_status. Fixes real duplicate message bug. Open since 7/26 with no maintainer review. Requesting review from @Teknium or team.