Reasoning skipped when using custom provider - API key availability check doesn't account for custom endpoints

[dguendisch]

Bug Description

When using reasoningProvider: 'custom' (e.g. pointing to a local proxy like LiteLLM), the reasoning pipeline is silently skipped because isAvailable() only checks for API keys of built-in providers (OpenAI, Anthropic, Gemini, Groq, Local). The custom provider is never evaluated, so reasoning is always marked as unavailable.

This affects users who use a local model for transcription (Whisper/Parakeet) combined with a cloud model for intelligence via a custom endpoint/proxy.

Steps to Reproduce

1. Set up a local STT model (e.g. Whisper)
2. Configure reasoning with provider custom, pointing to a proxy (e.g. LiteLLM)
3. Set a reasoning model (e.g. gemini-2.5-flash)
4. Perform a transcription

Expected: Transcribed text is processed by the reasoning model (filler word removal, etc.)
Actual: Reasoning is silently skipped. No call is made to the custom endpoint.

Debug Logs

[DEBUG][reasoning][renderer] REASONING_STORAGE_CHECK { useReasoning: true }
[DEBUG][reasoning][renderer] API_KEY_CHECK {
hasOpenAI: false,
hasAnthropic: false,
hasGemini: false,
hasGroq: false,
hasLocal: false
}
[DEBUG][reasoning][renderer] REASONING_AVAILABILITY { isAvailable: false, reasoningEnabled: true, finalDecision: false }
[DEBUG][reasoning][renderer] REASONING_CHECK {
useReasoning: false,
reasoningModel: 'gemini-2.5-flash',
reasoningProvider: 'custom',
agentName: 'OpenWhispr'
}

Root Cause

isReasoningAvailable() (in audioManager.js) calls ReasoningService.isAvailable(), which only checks for the presence of built-in provider API keys. The custom provider is not considered, so the check returns false even when the custom endpoint is fully configured.

Suggested Fix

isAvailable() should return true when reasoningProvider === 'custom' and a custom endpoint URL is configured.

Workaround

Set a dummy API key for any built-in provider (e.g. paste dummy into the OpenAI key field) to pass the gate check. The actual reasoning call still routes correctly through the custom provider.

I have verified that adding a dummy OpenAI key makes the custom provider reasoning work as expected.