Honor provider cooldown hints and add gateway OpenAI compat

[littlepenguin66]

Summary

• Problem:
  The provider cooldown path ignored shorter provider retry hints, and the gateway still lacked a complete OpenAI-compatible /v1/chat/completions + /v1/models surface that inherited shared runtime semantics.
• Why it matters:
  Ignoring provider rate-limit hints causes unnecessary wait or premature retries, and a provider-only gateway compat path drifts from shared turn/runtime behavior, usage reporting, and streaming semantics expected by OpenAI-compatible clients.
• What changed:
  Added provider rate-limit observation/parsing and threaded it through transport/failover/cooldown selection; made cooldown resolution honor provider hints when present; added the gateway OpenAI-compatible surface; routed accepted compat requests through the shared runtime path with usage propagation; disambiguated duplicate model ids; ensured streaming error termination with [DONE]; tightened Feishu/WeCom streaming request assertions.
• What did not change (scope boundary):
  Tool calling on the OpenAI-compatible gateway surface is still explicitly unsupported; channel entrypoints did not gain a new high-level fallback integration test for streaming-disabled transports.

Linked Issues

• Closes Rate-limit header observation for provider cooldown #983
• Closes OpenAI-compatible API surface in the gateway #996
• Related Extract ProviderTransport trait for testability #972
• Related Transport-level retry engine with exponential backoff #984

Change Type

• Bug fix
• Feature
• Refactor
• Documentation
• Security hardening
• CI / workflow / release

Touched Areas

• Kernel / policy / approvals
• Contracts / protocol / spec
• Daemon / CLI / install
• Providers / routing
• Tools
• Browser automation
• Channels / integrations
• ACP / conversation / session runtime
• Memory / context assembly
• Config / migration / onboarding
• Docs / contributor workflow
• CI / release / workflows

Risk Track

• Track A (routine / low-risk)
• Track B (higher-risk / policy-impacting)

If Track B, fill these in:

• Risk notes:
  This changes provider cooldown behavior, provider request parsing, shared turn/runtime wiring, gateway request validation, and streaming behavior.
• Rollout / guardrails:
  Gateway compat rejects unsupported tool fields and non-user-ending message sequences; duplicate model ids are surfaced as profile_id:model; package-level tests cover provider, gateway, and channel paths.
• Rollback path:
  Revert commit 6b617e3f (and follow-up doc commit if needed) to restore the pre-change provider cooldown and gateway behavior.

Validation

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• cargo test --workspace --locked
• cargo test --workspace --all-features --locked
• Relevant architecture / dep-graph / docs checks for touched areas
• Additional scenario, benchmark, or manual checks when behavior changed
• If this changes config/env fallback, limits, or defaults: include before/after behavior and regression coverage for explicit path, fallback path, and boundary values
• If tests mutate process-global env: document how state is restored or serialized

Commands and evidence:

cargo fmt --all -- --check
cargo test -p loongclaw-app --lib
cargo test -p loongclaw

All three commands passed locally after the final fix set.

Before/after notes:
- Provider cooldowns now prefer provider hints even when shorter than the previous policy floor.
- Gateway compat accepted requests now route through shared runtime semantics instead of direct provider execution.
- Duplicate configured model ids are exposed as stable `profile_id:model` aliases.
- Streaming error responses now terminate with `[DONE]`.
- OpenAI-compatible requests that do not end in a `user` message are rejected instead of silently bypassing the shared runtime path.

User-visible / Operator-visible Changes

• OpenAI-compatible clients can call /v1/models and /v1/chat/completions on the gateway and get shared-runtime behavior, usage propagation, stable duplicate-model aliases, and terminated SSE streams.
• Provider cooldown behavior now honors provider retry hints more precisely.

Failure Recovery

• Fast rollback or disable path:
  Revert the branch tip or disable use of the OpenAI-compatible gateway surface until the revert lands.
• Observable failure symptoms reviewers should watch for:
  Unexpected 400s for malformed/non-user-ending gateway chat payloads, incorrect model alias selection, missing usage in non-streaming compat responses, or channels no longer sending "stream": true in provider requests.

Reviewer Focus

• Provider hint parsing and cooldown precedence in crates/app/src/provider/rate_limit.rs and crates/app/src/provider/model_candidate_cooldown_runtime.rs.
• Shared-runtime routing and usage propagation in crates/daemon/src/gateway/openai_compat.rs, crates/app/src/agent_runtime.rs, crates/app/src/chat.rs, and crates/app/src/conversation/turn_coordinator.rs.
• Channel streaming assertions in crates/app/src/channel/feishu/webhook.rs, crates/app/src/channel/feishu/websocket.rs, and crates/app/src/channel/wecom.rs.

[gh-xj]

LoongClaw QA Review — PR #1233

Reviewed commit: 77b73ddd
Risk: high
Agent: ai-scientist

Findings

• high: Streamed OpenAI-compatible turn execution can drop streamed tool calls because the streaming parser only reconstructs text deltas and finish markers, not streamed tool-call chunks. crates/app/src/provider/contracts.rs:64
• medium: The non-streaming OpenAI-compatible gateway flattens propagated upstream 4xx/429 failures into HTTP 500, which breaks client-visible retry/auth semantics. crates/daemon/src/gateway/openai_compat.rs:229

Coverage

• Rust-specific review: applied (manual code-path review plus targeted Rust tests)
• Harness review: applied (repo-owned tests green; wrapper-path loongclaw-dev QA still lacks behavioral coverage)
• Adversarial challenge: applied (the participant_id QA build blocker did not reproduce under direct cargo build, so it remains a harness inconsistency rather than a confirmed branch defect)

Open Questions

• GitHub currently reports this PR as DIRTY; rebase is required before closure.
• Wrapper-path E2E proof through the OpenAI-compatible gateway is still missing and should be rerun after rebase from a clean QA lane.

Verdict

Blocking findings on the current head; rebase onto dev, fix the two runtime issues, and rerun boundary-aligned gateway QA before re-dispatch.