perf: 加速语音转写缓存命中

迁移旧版语音转写缓存，并将补迁失败降级为一次性失败状态。

按账号和消息标识优先命中兼容缓存，未命中时才读取音频并计算哈希；导出缓存命中后合并异步刷新知识索引。

补充迁移、缓存快速路径、批量语音读取和导出流程测试。

tests/unit/voice-pipeline.test.ts

@@ -3,6 +3,8 @@ import { tmpdir } from 'os'
 import { join } from 'path'
 import { afterAll, describe, expect, it, vi } from 'vitest'
 import { PcmAudioProcessor } from '../../src/main/voice-pipeline/audio-processor'
+import { VoicePipeline } from '../../src/main/voice-pipeline/voice-pipeline'
+import { voiceMessageIdentity } from '../../src/main/voice-pipeline/voice-message-identity'
 import { VoiceTaskScheduler } from '../../src/main/voice-pipeline/task-scheduler'
 import { SqliteTranscriptRepository } from '../../src/main/voice-pipeline/transcript-repository'
 import type { TranscriptRecord } from '../../src/main/voice-pipeline/types'
@@ -114,9 +116,13 @@ describe('voice task scheduling', () => {
           releaseFirst = resolve
         })
     )
-    const background = scheduler.schedule('background', async () => {
-      order.push('background')
-    }, { priority: 'background' })
+    const background = scheduler.schedule(
+      'background',
+      async () => {
+        order.push('background')
+      },
+      { priority: 'background' }
+    )
     const interactive = scheduler.schedule('interactive', async () => {
       order.push('interactive')
     })
@@ -134,7 +140,9 @@ describe('voice task scheduling', () => {
       'background',
       async (signal) => {
         order.push('background:start')
-        await new Promise<void>((resolve) => signal.addEventListener('abort', resolve, { once: true }))
+        await new Promise<void>((resolve) =>
+          signal.addEventListener('abort', resolve, { once: true })
+        )
         order.push('background:aborted')
         throw new DOMException('Recognition cancelled', 'AbortError')
       },
@@ -201,4 +209,155 @@ describe('transcript repository', () => {
     })
     repository.close()
   })
+
+  it('merges legacy transcripts idempotently without overwriting current records', () => {
+    const legacyPath = join(root, 'legacy-transcripts.sqlite')
+    const currentPath = join(root, 'current-transcripts.sqlite')
+    const legacy = new SqliteTranscriptRepository(legacyPath)
+    const current = new SqliteTranscriptRepository(currentPath)
+    const base: TranscriptRecord = {
+      accountId: 'account-a',
+      messageIdentity: 'message-1',
+      audioHash: 'audio-1',
+      processorVersion: 'processor-v1',
+      recognizerId: 'sensevoice',
+      modelVersion: 'model-v1',
+      modelFingerprint: 'fingerprint-a',
+      transcript: '旧缓存文字',
+      durationMs: 1200,
+      createdAt: 1,
+      updatedAt: 1
+    }
+    legacy.save(base)
+    legacy.save({ ...base, messageIdentity: 'message-2', audioHash: 'audio-2' })
+    current.save({ ...base, transcript: '当前缓存文字', updatedAt: 2 })
+    legacy.close()
+
+    expect(current.mergeFrom(legacyPath)).toBe(1)
+    expect(current.mergeFrom(legacyPath)).toBe(0)
+    expect(
+      current.find({
+        accountId: base.accountId,
+        messageIdentity: base.messageIdentity,
+        audioHash: base.audioHash,
+        processorVersion: base.processorVersion,
+        recognizerId: base.recognizerId,
+        modelVersion: base.modelVersion,
+        modelFingerprint: base.modelFingerprint
+      })?.transcript
+    ).toBe('当前缓存文字')
+    expect(current.findLatest('account-a', 'message-2')?.transcript).toBe('旧缓存文字')
+    current.close()
+  })
+})
+
+describe('voice pipeline cache lookup', () => {
+  it('returns a compatible message cache before reading or decoding audio', async () => {
+    const repository = new SqliteTranscriptRepository(join(root, 'fast-cache.sqlite'))
+    const reference = { sessionId: 'session', localId: 1, createTime: 2 }
+    repository.save({
+      accountId: 'account-a',
+      messageIdentity: voiceMessageIdentity(reference),
+      audioHash: 'audio-1',
+      processorVersion: 'processor-v1',
+      recognizerId: 'sensevoice',
+      modelVersion: 'model-v1',
+      modelFingerprint: 'fingerprint-a',
+      transcript: '快速命中',
+      durationMs: 900,
+      createdAt: 1,
+      updatedAt: 1
+    })
+    const resolve = vi.fn()
+    const decode = vi.fn()
+    const process = vi.fn()
+    const recognize = vi.fn()
+    const pipeline = new VoicePipeline(
+      { resolve },
+      { decode } as never,
+      { version: 'processor-v1', process },
+      {
+        metadata: {
+          recognizerId: 'sensevoice',
+          modelVersion: 'model-v1',
+          modelFingerprint: 'fingerprint-a'
+        },
+        recognize,
+        dispose: vi.fn()
+      },
+      repository
+    )
+
+    await expect(pipeline.run('account-a', reference)).resolves.toMatchObject({
+      transcript: '快速命中',
+      cached: true
+    })
+    expect(resolve).not.toHaveBeenCalled()
+    expect(decode).not.toHaveBeenCalled()
+    expect(process).not.toHaveBeenCalled()
+    expect(recognize).not.toHaveBeenCalled()
+    repository.close()
+  })
+
+  it('falls through when the cached processor version is incompatible', async () => {
+    const repository = new SqliteTranscriptRepository(join(root, 'version-cache.sqlite'))
+    const reference = { sessionId: 'session', localId: 1, createTime: 2 }
+    repository.save({
+      accountId: 'account-a',
+      messageIdentity: voiceMessageIdentity(reference),
+      audioHash: 'old-audio',
+      processorVersion: 'processor-v0',
+      recognizerId: 'sensevoice',
+      modelVersion: 'model-v1',
+      modelFingerprint: 'fingerprint-a',
+      transcript: '不兼容缓存',
+      durationMs: 900,
+      createdAt: 1,
+      updatedAt: 1
+    })
+    const resolve = vi.fn().mockResolvedValue({
+      data: Buffer.from('encoded'),
+      codec: 'silk',
+      sourceHash: 'new-audio'
+    })
+    const pipeline = new VoicePipeline(
+      { resolve },
+      {
+        decode: vi.fn().mockResolvedValue({
+          pcm: Buffer.from([1, 0]),
+          sampleRate: 16000,
+          channels: 1,
+          sourceHash: 'new-audio'
+        })
+      } as never,
+      {
+        version: 'processor-v1',
+        process: vi.fn().mockReturnValue({
+          samples: new Float32Array([0.1]),
+          sampleRate: 16000,
+          channels: 1,
+          sourceHash: 'new-audio',
+          processorVersion: 'processor-v1',
+          durationMs: 1
+        })
+      },
+      {
+        metadata: {
+          recognizerId: 'sensevoice',
+          modelVersion: 'model-v1',
+          modelFingerprint: 'fingerprint-a'
+        },
+        recognize: vi.fn().mockResolvedValue({ text: '新转写' }),
+        dispose: vi.fn()
+      },
+      repository
+    )
+
+    await expect(pipeline.run('account-a', reference)).resolves.toMatchObject({
+      transcript: '新转写',
+      cached: false
+    })
+    expect(resolve).toHaveBeenCalledOnce()
+    repository.close()
+  })
 })