fix(flow): drain trailing audio and harden chunked ASR pipeline

- P0/P1: tail-drain state machine in FlowContinuousCapture with converter
  flush; FlowSessionManager awaits drain before finalize
- P2: FlowCaptureTailDrain policy/tracker, pipeline diagnostics, unit tests
- P3: short final chunk merged re-transcription in ChunkedUtterancePipeline
- P4: UtteranceTranscriptStitcher composedSafely fallback; preview path parity
  in LiveDictationController

Co-authored-by: Rocky <hkgood@users.noreply.github.com>

OSGKeyboardTests/ChunkedUtterancePipelineTests.swift

@@ -87,6 +87,34 @@ final class ChunkedUtterancePipelineTests: XCTestCase {
         XCTAssertFalse(success.text.isEmpty)
         XCTAssertEqual(success.chunkWarnings.count, 1)
     }
+
+    func testPipelineRetranscribesShortFinalChunkWithPriorOverlap() async {
+        let config = FlowUtteranceChunkConfig(
+            maxChunkDurationSeconds: 0.05,
+            overlapDurationSeconds: 10,
+            pauseExtensionMaxSeconds: 0,
+            pauseRMSThreshold: 0.02,
+            minFinalChunkDurationSeconds: 0.05,
+            sampleRate: 1_000
+        )
+        let asr = ShortFinalMergeStubASR()
+        let pipeline = ChunkedUtterancePipeline(
+            asr: asr,
+            locale: Locale(identifier: "zh-Hans"),
+            config: config
+        )
+
+        let (stream, continuation) = AsyncStream<AudioBufferSnapshot>.makeStream()
+        continuation.yield(AudioBufferSnapshot(samples: [Float](repeating: 0.1, count: 80), sampleRate: 1_000))
+        continuation.yield(AudioBufferSnapshot(samples: [Float](repeating: 0.1, count: 20), sampleRate: 1_000))
+        continuation.finish()
+
+        let outcome = await pipeline.transcribe(stream: stream) { _ in }
+        guard case .success(let success) = outcome else {
+            return XCTFail("expected success, got \(outcome)")
+        }
+        XCTAssertTrue(success.text.contains("merged"))
+    }
 }
 
 private struct FailingSecondChunkASR: ASRService, @unchecked Sendable {
@@ -114,3 +142,32 @@ private struct FailingSecondChunkASR: ASRService, @unchecked Sendable {
         return .success("seg\(samples.count)")
     }
 }
+
+private struct ShortFinalMergeStubASR: ASRService, @unchecked Sendable {
+    private let callIndex = OSAllocatedUnfairLock(initialState: 0)
+
+    func transcribe(
+        stream: AsyncStream<AudioBufferSnapshot>,
+        locale: Locale
+    ) -> AsyncStream<ASREvent> {
+        AsyncStream { $0.finish() }
+    }
+
+    func cancel() {}
+
+    func transcribeChunk(samples: [Float], locale: Locale) async -> ASRChunkResult {
+        _ = locale
+        let current = callIndex.withLock { state in
+            let value = state
+            state += 1
+            return value
+        }
+        if current == 0 {
+            return .success("head")
+        }
+        if samples.count > 20 {
+            return .success("merged-tail")
+        }
+        return .success("short")
+    }
+}

OSGKeyboardTests/FlowCaptureTailDrainTests.swift

@@ -0,0 +1,51 @@
+// FlowCaptureTailDrainTests.swift
+// OSGKeyboardTests
+
+import XCTest
+@testable import OSGKeyboardShared
+
+final class FlowCaptureTailDrainTests: XCTestCase {
+
+    private let policy = FlowCaptureTailDrainPolicy(
+        silenceRMSThreshold: 0.02,
+        silenceDurationSeconds: 0.2,
+        maxDrainSeconds: 1.0
+    )
+
+    func testDrainFinishesAfterContinuousSilence() {
+        let tracker = FlowCaptureDrainTracker()
+        let start = Date().timeIntervalSince1970
+        tracker.beginDrain(now: start)
+
+        tracker.noteAudio(samples: [0.001, 0.001], policy: policy, now: start + 0.05)
+
+        let silentAt = start + 0.1
+        let decision = tracker.shouldFinish(policy: policy, now: silentAt + policy.silenceDurationSeconds)
+        XCTAssertTrue(decision.finished)
+        XCTAssertTrue(decision.endedBySilence)
+    }
+
+    func testDrainFinishesAtMaxDurationEvenWithoutSilence() {
+        let tracker = FlowCaptureDrainTracker()
+        let start = Date().timeIntervalSince1970
+        tracker.beginDrain(now: start)
+
+        tracker.noteAudio(samples: [0.5, 0.4], policy: policy, now: start + 0.05)
+        tracker.noteAudio(samples: [0.45, 0.42], policy: policy, now: start + 0.4)
+
+        let decision = tracker.shouldFinish(policy: policy, now: start + policy.maxDrainSeconds)
+        XCTAssertTrue(decision.finished)
+        XCTAssertFalse(decision.endedBySilence)
+    }
+
+    func testRMSDetectsAudibleSamples() {
+        XCTAssertGreaterThan(
+            FlowCaptureDrainTracker.rms(of: [0.2, 0.18, 0.15]),
+            policy.silenceRMSThreshold
+        )
+        XCTAssertLessThan(
+            FlowCaptureDrainTracker.rms(of: [0.001, 0.0005]),
+            policy.silenceRMSThreshold
+        )
+    }
+}

OSGKeyboardTests/UtteranceStreamChunkerTests.swift

@@ -63,4 +63,33 @@ final class UtteranceStreamChunkerTests: XCTestCase {
         XCTAssertGreaterThanOrEqual(received.count, 2)
         XCTAssertTrue(received.last?.isLast == true)
     }
+
+    func testFinalChunkIncludesLateArrivingTailSamples() async {
+        let config = FlowUtteranceChunkConfig(
+            firstChunkDurationSeconds: 0.5,
+            subsequentChunkDurationSeconds: 1.0,
+            overlapDurationSeconds: 0,
+            pauseExtensionMaxSeconds: 0,
+            pauseRMSThreshold: 0.02,
+            sampleRate: 1_000
+        )
+        let head = [Float](repeating: 0.05, count: 600)
+        let tail = [Float](repeating: 0.08, count: 250)
+
+        let (stream, continuation) = AsyncStream<AudioBufferSnapshot>.makeStream()
+        continuation.yield(AudioBufferSnapshot(samples: head, sampleRate: Double(config.sampleRate)))
+        continuation.yield(AudioBufferSnapshot(samples: tail, sampleRate: Double(config.sampleRate)))
+        continuation.finish()
+
+        var received: [UtteranceAudioChunk] = []
+        for await chunk in UtteranceStreamChunker.chunks(from: stream, config: config) {
+            received.append(chunk)
+        }
+
+        guard let last = received.last else {
+            return XCTFail("expected at least one chunk")
+        }
+        XCTAssertTrue(last.isLast)
+        XCTAssertGreaterThanOrEqual(last.samples.count, tail.count)
+    }
 }

OSGKeyboardTests/UtteranceTranscriptStitcherTests.swift

@@ -20,4 +20,24 @@ final class UtteranceTranscriptStitcherTests: XCTestCase {
         stitcher.append(index: 0, text: "第一段")
         XCTAssertEqual(stitcher.composed(), "第一段 第二段")
     }
+
+    func testComposedSafelyFallsBackWhenOverlapMergeShortensTooMuch() {
+        var stitcher = UtteranceTranscriptStitcher()
+        stitcher.append(index: 0, text: "今天天气很好我们")
+        stitcher.append(index: 1, text: "去公园")
+        let merged = stitcher.composed()
+        let safe = stitcher.composedSafely()
+        XCTAssertFalse(merged.isEmpty)
+        XCTAssertFalse(safe.isEmpty)
+        XCTAssertTrue(safe.contains("去公园"))
+    }
+
+    func testRemoveLastSegmentSupportsMergedTailRetranscription() {
+        var stitcher = UtteranceTranscriptStitcher()
+        stitcher.append(index: 0, text: "第一段")
+        stitcher.append(index: 1, text: "第二段")
+        stitcher.removeLastSegment()
+        stitcher.append(index: 1, text: "第二段合并")
+        XCTAssertEqual(stitcher.composed(), "第一段 第二段合并")
+    }
 }