fix(edge): stop short clips waiting a minute on the NAS and make the OpenAI STT fallback usable

A five-second phone recording took over a minute: stt-proxy waited up to
60 s for the self-hosted gateway, whose GPU endpoint was off and whose NAS CPU
Whisper needs 30-90 s per clip. With a direct provider configured the gateway
now gets 5 s plus the clip length (30 s cap).

The direct OpenAI fallback never produced a result. The production key held
characters that are not valid in an HTTP header, so every request threw while
being built; provider keys are now stripped of BOM/zero-width characters and a
still-invalid key counts as not configured. whisper-1 verbose_json reports the
language by name, which the result contract rejected; names now map to codes.
Fail-closed responses list each provider's failure (status or error class,
no secrets) so an outage can be diagnosed without log access.

server/supabase/functions/_shared/stt-contract.ts

@@ -1,6 +1,47 @@
 export const MAX_STT_AUDIO_BYTES = 25 * 1024 * 1024
 export const STT_PROVIDER_TIMEOUT_MS = 60_000
 
+/** 16 kHz 16-bit mono PCM — the most common upload; compressed audio only shortens the estimate. */
+const PCM_BYTES_PER_SECOND = 32_000
+const GATEWAY_BASE_MS = 5_000
+const GATEWAY_MAX_WITH_FALLBACK_MS = 30_000
+
+/**
+ * How long to wait for the self-hosted gateway before a direct provider takes over.
+ *
+ * The gateway tries a GPU box that may be off and a CPU-only NAS Whisper that
+ * needed 50–90 s for a 3 s clip (2 cores under load, 2026-09-27). Waiting the
+ * full provider timeout turned a short recording into a minute-long wait, and
+ * the phone gave up to its tiny local model. With a direct provider configured,
+ * give the gateway 5 s plus the clip length (enough for a GPU, far too little
+ * for the NAS) and fall through; without one, keep the full timeout.
+ */
+const LANGUAGE_NAME_TO_CODE: Readonly<Record<string, string>> = {
+  korean: 'ko', english: 'en', japanese: 'ja', chinese: 'zh', spanish: 'es', french: 'fr',
+  german: 'de', portuguese: 'pt', russian: 'ru', vietnamese: 'vi', thai: 'th', italian: 'it',
+  dutch: 'nl', indonesian: 'id', turkish: 'tr', arabic: 'ar', hindi: 'hi', polish: 'pl',
+}
+
+/**
+ * Whisper verbose_json (OpenAI, Groq) reports the language by name ("korean"), while the
+ * STT contract carries a code. A name used to fail normalizeSttResult, so every direct
+ * OpenAI/Groq fallback was thrown away as an invalid result. Accept a code, map a name,
+ * otherwise use what the client asked for.
+ */
+export function providerLanguageCode(reported: unknown, requested: string): string {
+  const fallback = requested === 'auto' || requested === 'multi' ? 'und' : requested
+  if (typeof reported !== 'string') return fallback
+  const value = reported.trim().toLowerCase()
+  if (/^[a-z]{2,3}(?:-[a-z0-9]{2,8})?$/.test(value)) return value
+  return LANGUAGE_NAME_TO_CODE[value] ?? fallback
+}
+
+export function gatewayDeadlineMs(audioBytes: number, hasDirectFallback: boolean): number {
+  if (!hasDirectFallback) return STT_PROVIDER_TIMEOUT_MS
+  const estimatedSeconds = Math.max(0, audioBytes) / PCM_BYTES_PER_SECOND
+  return Math.min(GATEWAY_MAX_WITH_FALLBACK_MS, Math.round(GATEWAY_BASE_MS + estimatedSeconds * 1_000))
+}
+
 export interface NormalizedSttResult {
   transcript: string
   confidence: number