fix(asr): keep recordings when streaming recognition fails or stalls

- Keep recording after a streaming ASR error and retry the captured PCM
  through batch ASR before reporting a failure.
- Treat an expired ASR wait as a stall so batch fallback still runs.
- Prime the managed Volcengine stream on mic touch and adopt it only while
  fresh and open.
- Auto-scroll the live transcript to the newest recognized text.

OSGKeyboard/Services/FlowSessionManager.swift

@@ -2109,6 +2109,13 @@ final class FlowSessionManager: ObservableObject {
               !hasUnacknowledgedTerminalResult() else {
             return
         }
+        bindSessionASRIfNeeded()
+        if let cloudASR = sessionASR as? CloudASRService {
+            // Same locale source as `beginUtterance`, so the prime is adoptable.
+            cloudASR.primeUtteranceStreaming(
+                locale: SpeechLocaleResolver.resolve(store.localeId)
+            )
+        }
         startAudioPrime(id: command.utteranceId, origin: "micTouch")
     }
 
@@ -2171,6 +2178,7 @@ final class FlowSessionManager: ObservableObject {
               !isUtteranceProcessing else {
             return
         }
+        (sessionASR as? CloudASRService)?.cancelPrimedUtteranceStreaming()
         audioPrimeCancellationRequested = true
         // `capture.start()` is not cancellation-safe. Never stop it mid-start;
         // the completion waiter performs cleanup unless a real utterance adopts
@@ -2613,11 +2621,14 @@ final class FlowSessionManager: ObservableObject {
                     ]
                         .map { $0.trimmingCharacters(in: .whitespacesAndNewlines) }
                         .first(where: { !$0.isEmpty })
+                    // Without a partial, keep recording: capture is independent
+                    // of ASR, and finalize retries the full PCM through batch
+                    // ASR (`recognitionFailed`) before reporting the error.
                     if let recovery {
                         manager.lastFinal = recovery
                         manager.debug("asr error recovered via partial len=\(recovery.count)")
                     } else if manager.isUtteranceRecording {
-                        manager.failUtterance(message: message, kind: .asrFailed)
+                        manager.debug("asr error while recording — deferring to batch fallback")
                     }
                 case .cancelled:
                     // A provider-side cancellation while the mic is still open
@@ -2935,8 +2946,16 @@ final class FlowSessionManager: ObservableObject {
             try? await Task.sleep(nanoseconds: 100_000_000)
         }
 
+        // The recognizer never reported back within the wait: treat it as a
+        // failure so batch ASR can still recover the captured audio.
+        let asrStalled = lastFinal.isEmpty
+            && asrCompletedGeneration != finalizeGeneration
+            && !Task.isCancelled
         if lastFinal.isEmpty {
-            FlowDiagnostics.log("ASR wait elapsed — cancelling ASR task and using best available transcript")
+            FlowDiagnostics.log(
+                "ASR wait elapsed — cancelling ASR task and using best available transcript "
+                    + "stalled=\(asrStalled ? 1 : 0)"
+            )
             asrTask?.cancel()
             Task { await chunkedPipeline?.cancel() }
         }
@@ -2973,11 +2992,13 @@ final class FlowSessionManager: ObservableObject {
         let wantsBatchFallback = UtteranceBatchFallbackPolicy.shouldRunBatchFallback(
             stitchedFinal: lastFinal,
             partialSnapshot: bestPartialSnapshot,
-            recognitionFailed: asrFailureMessage != nil
+            recognitionFailed: asrFailureMessage != nil,
+            recognitionStalled: asrStalled
         )
         FlowTrace.pipeline(
             "batchFallback.decision",
-            "wanted=\(wantsBatchFallback ? 1 : 0) pcmSamples=\(utterancePCMSamples.count) "
+            "wanted=\(wantsBatchFallback ? 1 : 0) stalled=\(asrStalled ? 1 : 0) "
+                + "pcmSamples=\(utterancePCMSamples.count) "
                 + "pcmRms=\(FlowTrace.rms(utterancePCMSamples)) "
                 + "stitchedLen=\(lastFinal.count) partialLen=\(bestPartialSnapshot.count) "
                 + "resolvedLen=\(text.count)"