[Bug] AI chat failures discard the provider's error; only "The AI service could not complete this request" is visible #441

Version: v0.9.2 (b8e4e66); the logging is unchanged on main (b1a9898)

When the model provider rejects a chat request, nothing in Windshift shows why:

• the chat panel says "The AI service could not complete this request";
• respondLLMError in internal/handlers/ai_helpers.go logs only error_type: *fmt.wrapError;
• agent_runs.error is the fixed string "Agent chat execution failed";
• the connection test passes, because it sends a one-token prompt with no tools, temperature or thinking options, so it does not exercise what the chat sends.

In our case the cause was Anthropic rejecting temperature on a Claude 5.5 model (fixed by dc618c0, thanks). Finding that took a capture proxy between Windshift and the provider, because the message was dropped at every layer. The daily briefing scheduler, by contrast, logs the full provider error (briefing generation failed ... 400 Bad Request ... "temperature is deprecated for this model"), which is how we eventually confirmed it.

Requests

1. Log the wrapped error text for chat and agent runs, as the briefing job already does. Provider messages do not contain secrets.
2. Store it on the run (agent_runs.error) and show it in the admin diagnostics, or at least in the chat panel for admins.
3. Make the connection test send the same parameter set the chat uses (temperature, a trivial tool, the configured reasoning option) so it fails when the chat would.