fix(vision,gateway): wall-clock timeout for vision calls + 502/503/504 error hint

[daniel-rudaev]

What

Two small resilience/UX fixes for provider degradation:

1. Bound vision calls with a wall-clock timeout (tools/vision_tools.py).
   vision_analyze_tool awaited async_call_llm with only the per-attempt SDK/transport timeout. When a provider degrades — a 503 storm, or a stalled socket the transport read-timeout never trips — the await can hang far past the configured timeout. Vision runs inside the agent turn, so a hung call blocks the entire conversation and the user gets no response at all. Fix: wrap the call in asyncio.wait_for with a wall-clock bound (auxiliary.vision.hard_timeout, default = per-attempt timeout + 60s) and return a clear "provider taking too long, try again" message on timeout, instead of hanging the turn silently.

2. Add a 502/503/504 hint to the agent error handler (gateway/run.py).
   The per-status hint ladder covered 401/402/429/529/400/500 but not the common 502/503/504 "provider temporarily unavailable" class, so a transient provider 5xx surfaced as a bare "Sorry, I encountered an error" with no actionable guidance. Adds a 5xx branch telling the user it's transient and to retry.

Why

A turn that exhausts providers/retries — or stalls on a slow vision provider — should never leave the user in silence, and a single stalled image-analysis turn should not block the whole conversation.

Notes

• Minimal + backward-compatible: hard_timeout defaults to timeout + 60s when unset, so existing configs are unchanged.
• Validated: the wall-clock guard fires and returns the timeout message through the real vision_analyze_tool path; both files compile.

[teknium1]

Thanks for addressing two live resilience gaps. Current main still directly awaits async_call_llm in tools/vision_tools.py:1253, and the gateway hint ladder in gateway/run.py:12258-12296 still omits 502/503/504.

Problems

• The new auxiliary.vision.hard_timeout surface is only read by the PR. It is absent from the defaults in hermes_cli/config.py:1577-1585 and the documented auxiliary reference at website/docs/user-guide/configuration.md:998-1009.
• The diff adds no regression coverage. Existing vision config coverage in tests/tools/test_vision_tools.py:424-481 covers timeout and temperature, not a stalled coroutine; no gateway test covers this status-hint branch.

Suggested changes

• Register and document hard_timeout alongside the existing vision timeout settings.
• Add hermetic tests for a never-completing vision call and for 502/503/504 gateway hints.

This is an automated hermes-sweeper review.

[teknium1]

hard_timeout is a new user-facing config key, but this diff does not register it in DEFAULT_CONFIG or document it. Please add the default and user-facing reference alongside the existing auxiliary.vision.timeout settings so it is discoverable and has stable config semantics.