feat: switch embeddings from Voyage AI to Google Gemini text-embedding-004

- embedText: call Gemini REST API (768-dim) instead of Voyage (1024-dim)
- Migration: drop+recreate incidents.embedding as vector(768), update
  match_incidents function, swap VOYAGE_API_KEY → GOOGLE_AI_API_KEY in app_settings
- Settings UI: relabel to "Google AI API Key (Embeddings)"
- All call sites updated (incidents POST, similar GET)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CPf5Rc8QPx6V8KLEEgfKEQ

app/api/incidents/[id]/similar/route.ts

@@ -18,7 +18,7 @@ export async function GET(
   if (!profile || !['hse', 'admin'].includes(profile.role))
     return NextResponse.json({ error: 'Forbidden' }, { status: 403 })
 
-  const voyageKey = await getApiKey(supabase, 'VOYAGE_API_KEY')
+  const googleAiKey = await getApiKey(supabase, 'GOOGLE_AI_API_KEY')
 
   const { data: incident } = await supabase
     .from('incidents')
@@ -34,7 +34,7 @@ export async function GET(
     if (inc.embedding) {
       embeddingVec = JSON.parse(inc.embedding) as number[]
     } else {
-      embeddingVec = await embedText(inc.description, voyageKey)
+      embeddingVec = await embedText(inc.description, googleAiKey)
       // Closed incidents are locked at the DB level — the trigger would reject
       // this backfill. The vector still serves the similarity query below.
       if (inc.status !== 'closed') {