feat(config): measured chat budget (90,000) and upload chat model (gemma4-26b)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

internal/config/uploads.go

@@ -33,9 +33,15 @@ type UploadConfig struct {
 	FFprobeBin   string
 	// MaxUploadedTranscriptChars bounds the transcript sent to a chat model for an
 	// UPLOADED recording. The 18,000-char default for captions is ~22 minutes of
-	// speech; an hour of meeting is ~50,000 chars. A guess until measured against
+	// speech; an hour of meeting is ~50,000 chars. MEASURED 2026-10-01 against the
-	// the models chat offers.
+	// real models: iguana/gemma4-26b answers correctly to 90,000 chars but silently
+	// drops the head of a 120,000-char transcript, so 90,000 is the ceiling of the
+	// best model, not a round number.
 	MaxUploadedTranscriptChars int
+	// UploadChatModel is the model chat over an uploaded recording opens with. It is
+	// NOT the summarizer default: koala/phi4-mini has an 8,192-token context and
+	// returns HTTP 400 on an hour of speech (~13,000 tokens).
+	UploadChatModel string
 }
 
 const (
@@ -44,7 +50,8 @@ const (
 	defaultTranscribeModel         = "iguana/kb-whisper"
 	defaultTranscribeFallbackModel = "berget/kb-whisper"
 	defaultTranscribeTimeout       = 30 * time.Minute
-	defaultMaxUploadedChars        = 120000
+	defaultMaxUploadedChars        = 90000
+	defaultUploadChatModel         = "iguana/gemma4-26b"
 )
 
 func loadUploadConfig() (UploadConfig, error) {
@@ -54,6 +61,7 @@ func loadUploadConfig() (UploadConfig, error) {
 		UploadTmpDir:            envOr("TAPIR_UPLOAD_TMP_DIR", os.TempDir()),
 		FFmpegBin:               envOr("TAPIR_FFMPEG_BIN", "ffmpeg"),
 		FFprobeBin:              envOr("TAPIR_FFPROBE_BIN", "ffprobe"),
+		UploadChatModel:         envOr("TAPIR_UPLOAD_CHAT_MODEL", defaultUploadChatModel),
 	}
 
 	if v := os.Getenv("TAPIR_UPLOADS_ENABLED"); v != "" {