diff --git a/apps/api/src/modules/import/magic-bytes.ts b/apps/api/src/modules/import/magic-bytes.ts
new file mode 100644
index 0000000..bf79961
--- /dev/null
+++ b/apps/api/src/modules/import/magic-bytes.ts
@@ -0,0 +1,49 @@
+// 手写魔数签名表（Plan 3 Task 4，spec §11.1、计划"Tech Stack：零新增依赖"）：
+// 常见六类媒体格式判别 + SVG 文本特征检出。刻意不引 file-type 依赖——导入管道的
+// 安全决策只须回答两问："是什么格式"与"是否 SVG（隔离）"，一张可审计的小签名表
+// 足够，且避免为一次头部读取引入整个 MIME 数据库。
+//
+// 判别语义（detectMediaType）：
+// - 命中二进制签名 → {type: MIME, svg: false}（PNG/JPEG/GIF87a/GIF89a/WEBP/PDF）；
+// - 无二进制命中但前 1KB 文本窗口含 "<svg"（大小写不敏感）→ {type: 'image/svg+xml',
+//   svg: true}——调用方据此 status='quarantined'（spec §11.1：SVG 可携带脚本与外部
+//   引用，不进产物语义）；
+// - 均未命中 → null（未知签名——调用方 status='unsupported' + issue，不猜格式）。
+//
+// 顺序纪律：二进制签名先于 SVG 文本扫描——PNG/JPEG 正例即便尾部字节碰巧含 "<svg"
+// 片段也按二进制格式判（文本窗口只对非二进制文件有意义）。窗口 1KB 是钉死的折中：
+// 覆盖常规 "<?xml …?><svg" 前导（声明/注释/doctype），不给深藏脚本的巨型前缀放行。
+//
+// Buffer 读取口径：latin1 单字节解码后小写化做包含检测——不做 UTF-8 多字节解码，
+// 避免截断的字节序列被解码成意外字符。
+/** 二进制签名表（前缀匹配；WEBP 的 RIFF+WEBP@8 双段结构在表外单查） */
+const SIGNATURES: ReadonlyArray<{ mime: string; bytes: Buffer }> = Object.freeze([
+  { mime: 'image/png', bytes: Buffer.from([0x89, 0x50, 0x4e, 0x47]) },
+  { mime: 'image/jpeg', bytes: Buffer.from([0xff, 0xd8, 0xff]) },
+  { mime: 'image/gif', bytes: Buffer.from('GIF87a', 'latin1') },
+  { mime: 'image/gif', bytes: Buffer.from('GIF89a', 'latin1') },
+  { mime: 'application/pdf', bytes: Buffer.from('%PDF', 'latin1') },
+])
+
+const RIFF = Buffer.from('RIFF', 'latin1')
+const WEBP = Buffer.from('WEBP', 'latin1')
+const SVG_WINDOW_BYTES = 1024
+
+/**
+ * 魔数检测：读文件头部字节判别媒体类型。返回 null = 未知签名（调用方按
+ * unsupported 处置）；svg=true 仅出现在 type='image/svg+xml'（文本特征命中）。
+ */
+export function detectMediaType(buf: Buffer): { type: string; svg: boolean } | null {
+  for (const sig of SIGNATURES) {
+    if (buf.length >= sig.bytes.length && buf.subarray(0, sig.bytes.length).equals(sig.bytes)) {
+      return { type: sig.mime, svg: false }
+    }
+  }
+  // WEBP：RIFF 容器头（4B 长度）+ 'WEBP' Magic（@8）——RIFF 家族其他成员（WAV/AVI）不误报
+  if (buf.length >= 12 && buf.subarray(0, 4).equals(RIFF) && buf.subarray(8, 12).equals(WEBP)) {
+    return { type: 'image/webp', svg: false }
+  }
+  const head = buf.subarray(0, Math.min(buf.length, SVG_WINDOW_BYTES)).toString('latin1').toLowerCase()
+  if (head.includes('<svg')) return { type: 'image/svg+xml', svg: true }
+  return null
+}
diff --git a/apps/api/src/modules/import/media-writer.ts b/apps/api/src/modules/import/media-writer.ts
new file mode 100644
index 0000000..62c42cd
--- /dev/null
+++ b/apps/api/src/modules/import/media-writer.ts
@@ -0,0 +1,617 @@
+// 媒体登记写入器（Plan 3 Task 4，spec §5.5/§11.1、前置事实 7/11、评审 Issue 1/3）：
+// 消费适配器事件流的**双事件**面——attachment 行建资产（media_assets）、内容行提取引用
+// （media_usages）——媒体域的领域 owner（Task 6 分派规则：content-writer 对 attachment 行
+// 只留 source_records 足迹，media_assets 的唯一写入方在此）。
+//
+// 三级对拍（前置事实 7 钉死，真实数据实证：媒体目录扁平 GUID 命名、与 WXR
+// attachment_url 零精确命中，唯一映射在旧 cms 库侧车）：
+// ① uploads 相对路径清单精确命中：attachment_url 掐掉
+//    '…/wp-content/uploads/' 前缀 → manifest[path]（清单 = media_bundle 快照 metadata.
+//    manifest，{path, bytes, sha256}）；
+// ② 侧车命中：manifest miss → sidecarMap[source_url] → filePath → 清单实体（精确路径
+//    miss 时按**文件名**候选——真实侧车 file_path 是扁平文件名而清单可含子目录；候选
+//    多条 = 歧义，不自动选、全部列 issue、不插行，spec §5.5）；
+// ③ missing：两级皆 miss → 只进报告（**不插 media_assets 行**——sha256 NOT NULL 无清单
+//    实体可取，行插不进是库层防线不是巧合）。
+// 命中行取值：public_path = 旧 uploads 路径（保旧 URL 语义）、sha256/byte_size 从清单、
+// storage_path = bundle 副本相对路径 + 清单 path（绝对副本 = metadata.storageRoot + 此列）。
+//
+// 魔数安全校验（spec §11.1）：命中实体读存储副本头部实际字节（detectMediaType）——
+// - SVG 检出（文本特征）→ status='quarantined'：行存在（登记可审计）但不进产物语义，
+//   其引用不建 media_usages（usage 是发布面消费的引用关系，隔离资产不得经此进产物）；
+// - 未知签名 → status='unsupported' + issue（不猜格式）；
+// - 声明 MIME（清单/WXR 均无 MIME——声明侧由 URL 扩展名推断）与检测不符 → issue
+//   （仍登记 ready：内容真实、扩展名说谎是数据质量问题不是安全风险）。
+//
+// usage 时序（设计决定，WXR 文档序钉死）：items 先于 attachment——引用先行、资产后到，
+// 故 usage 建立分两阶段：内容行到达时提取引用并落修订锚点（source_records 六列身份 →
+// entry_locales.current_revision_id——**依赖 content-writer 先处理同事件**，Task 6 编排
+// 顺序契约），资产解析统一推迟到 finish() 对 media_assets 反查（id 反查 source_id /
+// URL 反查 source_url→public_path）——两种文档序终态一致。引用无对应资产且属本站媒体域
+// （id 引用或 uploads 路径 URL）→ mediaDanglingRefs 计数 + issue（§7.6 禁静默）；外域
+// 热链 URL 不属本域对拍范围，不计悬空（无清单可比对，静默越过）。
+//
+// 幂等：media_assets 唯一 (site, source_url) onConflictDoNothing（重跑零增行）；
+// media_usages 复合主键 (asset, revision, field_path) onConflictDoNothing——field_path
+// 由提取序确定（featured/gallery 序号/正文 img 序号），重跑同流同值。事务边界：全部
+// 单语句原子操作（ctx.db 直连约定，无跨语句不变量——与 term-writer 同纪律）。
+import { open, readFile } from 'node:fs/promises'
+import { join } from 'node:path'
+import { and, eq } from 'drizzle-orm'
+import type { AdapterEvent } from '@sf/adapter-wp'
+import type { UniversalRow } from '@sf/contracts'
+import { entryLocales, mediaAssets, mediaUsages, sourceRecords } from '../../db/schema.js'
+import { isUniqueViolation, type MediaManifestEntry } from '../source/register.service.js'
+import { detectMediaType } from './magic-bytes.js'
+import type { ImportDb, ImportJobContext, ImportPipeline, ImportPipelineOutcome } from './import.service.js'
+
+const UPLOADS_MARKER = '/wp-content/uploads/'
+
+/** 未知签名的 detected_mime 落值（列 NOT NULL——不猜格式但要有确定值） */
+const UNKNOWN_DETECTED_MIME = 'application/octet-stream'
+
+/** 魔数读取的头部字节（SVG 文本窗口 1KB + 余量） */
+const HEAD_BYTES = 4096
+
+/** URL 扩展名 → 声明 MIME（声明侧推断表——与 detectMediaType 的检出表独立） */
+const EXT_MIME: Readonly<Record<string, string>> = Object.freeze({
+  png: 'image/png',
+  jpg: 'image/jpeg',
+  jpeg: 'image/jpeg',
+  gif: 'image/gif',
+  webp: 'image/webp',
+  pdf: 'application/pdf',
+  svg: 'image/svg+xml',
+})
+
+/**
+ * 可分派的媒体写入器句柄（Task 6 编排形状，与 term/content 同构——但 finish 是异步：
+ * usage 解析推迟到流终，见模块头"usage 时序"）。句柄非可重入：一个句柄只服务一遍流。
+ */
+export interface MediaEventHandler {
+  /** 单事件分派（term/unsupported/error 为 no-op——无媒体信息） */
+  handleEvent(ev: AdapterEvent): Promise<void>
+  /** 流终：解析悬置引用 → media_usages / 悬空报告，取回累计产出（调用后句柄弃用） */
+  finish(): Promise<ImportPipelineOutcome>
+}
+
+/**
+ * 本写入器的 counts 键与归属（**counts 键是跨写入器的扁平全局命名空间**——import.service
+ * 按返回值覆盖累加器同名键合并，各写入器必须各持独占键名，见 ImportPipelineOutcome.counts
+ * 的键契约注记）。本写入器独占六键（media 前缀，**勿复用 content-writer 的 'skipped'**）：
+ * - mediaRegistered：本运行**新插**的 media_assets 行数（含 quarantined/unsupported——
+ *   行已登记；幂等重跑为 0）；
+ * - mediaFromSidecar：其中经级②侧车解析的行数（mediaRegistered 的子集）；
+ * - mediaMissing：级③ missing 的 attachment 数（按事件计——重跑如实再现）；
+ * - mediaQuarantined：本运行检出 SVG 的 attachment 数（按事件计——重跑如实再现，
+ *   与 mediaRegistered 在重跑时不再严格子集：行幂等而检出如实）；
+ * - mediaDanglingRefs：内容引用未解析到资产的条数（按引用计——重跑如实再现）；
+ * - mediaAmbiguous：多候选歧义的 attachment 数（按事件计，不插行）。
+ * issues：missing/歧义/隔离/未知签名/MIME 不符/悬空引用/副本读取失败/public_path 唯一键
+ * 冲突——全部显式报告（§7.6 禁静默）。
+ */
+interface MediaWriterCounts {
+  mediaRegistered: number
+  mediaFromSidecar: number
+  mediaMissing: number
+  mediaQuarantined: number
+  mediaDanglingRefs: number
+  mediaAmbiguous: number
+}
+
+/** 悬置引用（内容行提取、流终解析）：锚点 + 引用键两半 */
+interface PendingRef {
+  /** 引用方修订（content-writer 落的 current_revision_id——usage 的不可变锚） */
+  revisionId: string
+  /** 修订数据域内的字段位（同流重跑稳定——复合主键的 field_path 侧） */
+  fieldPath: string
+  kind: 'id' | 'url'
+  /** id 引用 = attachment post id；URL 引用 = 引文原串 */
+  value: string
+  /** URL 引用属本站 uploads 域时的公开路径形态（悬空裁决域）；id 引用恒属本域 */
+  uploadsPublicPath: string | null
+  rawLocation: string
+}
+
+/** 侧车 JSON 条目形状（register --media-map 已在登记期校验，此处只做读取防御） */
+interface SidecarEntry {
+  sourceUrl: string
+  filePath: string
+}
+
+export function createMediaEventHandler(identity: {
+  db: ImportDb
+  siteId: string
+  connectionId: string
+  snapshotId: string
+  /** 引用行 locale 校验基准（与 content-writer 同拒收口径——不双报） */
+  defaultLocale: string
+  /** 存储副本绝对根（media_bundle 快照 metadata.storageRoot——登记期事实） */
+  storageRoot: string
+  /** media_bundle 快照的 storagePath（相对 storageRoot；直传清单场景传 ''） */
+  mediaBundleStoragePath: string
+  /** 媒体清单（media_bundle metadata.manifest；缺席 = 空表 → 全部走级③） */
+  mediaManifest: MediaManifestEntry[]
+  /** 侧车映射（manual_url_list：source_url → file_path） */
+  sidecarMap: Map<string, string>
+}): MediaEventHandler {
+  const { db, siteId } = identity
+  const counts: MediaWriterCounts & Record<string, number> = {
+    mediaRegistered: 0,
+    mediaFromSidecar: 0,
+    mediaMissing: 0,
+    mediaQuarantined: 0,
+    mediaDanglingRefs: 0,
+    mediaAmbiguous: 0,
+  }
+  const issues: string[] = []
+  const pendingRefs: PendingRef[] = []
+
+  // 清单索引：精确路径 + 文件名候选（级②的侧车 file_path 是扁平名，清单可含子目录）
+  const manifestByPath = new Map<string, MediaManifestEntry>()
+  const manifestByBasename = new Map<string, MediaManifestEntry[]>()
+  for (const entry of identity.mediaManifest) {
+    manifestByPath.set(entry.path, entry)
+    const base = basenameOf(entry.path)
+    const bucket = manifestByBasename.get(base)
+    if (bucket === undefined) manifestByBasename.set(base, [entry])
+    else bucket.push(entry)
+  }
+
+  function basenameOf(path: string): string {
+    const idx = path.lastIndexOf('/')
+    return idx === -1 ? path : path.slice(idx + 1)
+  }
+
+  /** 扩展名推断声明 MIME（无扩展名/表外扩展 → null：列可空，无声明则无对拍） */
+  function declaredMimeOf(pathLike: string): string | null {
+    const m = /\.([a-z0-9]+)$/i.exec(pathLike)
+    if (m === null) return null
+    const hit = EXT_MIME[m[1]!.toLowerCase()]
+    return hit ?? null
+  }
+
+  /**
+   * attachment_url → 旧 uploads 相对路径。无 uploads 标记的 URL（外站迁移源）级①无从
+   * 掐前缀：relPath=null（级①跳过），source/public 路径以 URL pathname 兜底。
+   * 注：不做百分号解码——清单 path 是字面文件名，编码差异按 miss 如实进报告。
+   */
+  function uploadsRelOf(fileUrl: string): { relPath: string | null; pathname: string } {
+    const idx = fileUrl.indexOf(UPLOADS_MARKER)
+    if (idx !== -1) {
+      const rel = fileUrl.slice(idx + UPLOADS_MARKER.length)
+      if (rel !== '') return { relPath: rel, pathname: fileUrl.slice(idx) }
+    }
+    try {
+      return { relPath: null, pathname: new URL(fileUrl).pathname }
+    } catch {
+      // 非 URL 形态的原始串： pathname 兜底保留原文（可审计），级①不可用
+      return { relPath: null, pathname: fileUrl.startsWith('/') ? fileUrl : `/${fileUrl}` }
+    }
+  }
+
+  /** 存储副本头部读取（魔数只需前若干字节——171MB 级文件不做全量读） */
+  async function readHead(abs: string): Promise<Buffer> {
+    const handle = await open(abs, 'r')
+    try {
+      const buf = Buffer.alloc(HEAD_BYTES)
+      const { bytesRead } = await handle.read(buf, 0, HEAD_BYTES, 0)
+      return buf.subarray(0, bytesRead)
+    } finally {
+      await handle.close()
+    }
+  }
+
+  /**
+   * attachment 行主路径：三级解析 → 魔数 → 插 media_assets。
+   * locale 拒收口径与 content-writer 一致（该行已被 content-writer 拒收并记 issue，
+   * 此处静默跳过不双报——attachment 行的媒体域处理同样以 locale 合法为前提）。
+   */
+  async function handleAttachmentRow(row: UniversalRow): Promise<void> {
+    if (row.locale !== identity.defaultLocale) return
+    const fileUrl = typeof row.knownFields.attachmentUrl === 'string' ? row.knownFields.attachmentUrl : ''
+    if (fileUrl === '') {
+      counts.mediaMissing++
+      issues.push(
+        `attachment 缺 attachment_url（${row.sourceType}/${row.sourceId}，${row.rawLocation}）——三级对拍无从下手，按 missing 报告`,
+      )
+      return
+    }
+    const { relPath, pathname } = uploadsRelOf(fileUrl)
+
+    // 级①：uploads 相对路径清单精确命中
+    let entry: MediaManifestEntry | null = relPath === null ? null : manifestByPath.get(relPath) ?? null
+    let fromSidecar = false
+    if (entry === null) {
+      // 级②：侧车 source_url → file_path → 清单实体（精确路径 → 文件名候选）
+      const filePath = identity.sidecarMap.get(fileUrl)
+      if (filePath !== undefined) {
+        const exact = manifestByPath.get(filePath)
+        if (exact !== undefined) {
+          entry = exact
+          fromSidecar = true
+        } else {
+          const candidates = manifestByBasename.get(basenameOf(filePath)) ?? []
+          if (candidates.length > 1) {
+            // 歧义不自动选（spec §5.5）：全部候选列报告，不插行
+            counts.mediaAmbiguous++
+            issues.push(
+              `媒体多候选歧义（${row.sourceType}/${row.sourceId}，${fileUrl}）：侧车 file_path=${filePath} 按文件名命中 ${candidates.length} 条清单实体——[${candidates.map((c) => c.path).join('、')}]——不自动选择，未登记`,
+            )
+            return
+          }
+          if (candidates.length === 1) {
+            entry = candidates[0]!
+            fromSidecar = true
+          }
+        }
+      }
+    }
+    if (entry === null) {
+      // 级③：missing——只进报告不插行（sha256 NOT NULL：无清单实体插不进，库层防线）
+      counts.mediaMissing++
+      issues.push(
+        `媒体 missing（${row.sourceType}/${row.sourceId}，${fileUrl}）：清单与侧车均未命中——不插 media_assets 行（sha256 无来源）`,
+      )
+      return
+    }
+
+    // 魔数：读存储副本实际字节（storageRoot + bundle 相对路径 + 清单 path）
+    const storagePath =
+      identity.mediaBundleStoragePath === ''
+        ? entry.path
+        : `${identity.mediaBundleStoragePath}/${entry.path}`
+    let head: Buffer
+    try {
+      head = await readHead(join(identity.storageRoot, storagePath))
+    } catch (err) {
+      issues.push(
+        `媒体副本读取失败（${row.sourceType}/${row.sourceId}，${fileUrl} → ${storagePath}）：${err instanceof Error ? err.message : String(err)}——清单声称存在而副本缺席，未登记`,
+      )
+      return
+    }
+    const detected = detectMediaType(head)
+    const publicPath = relPath === null ? pathname : `${UPLOADS_MARKER}${relPath}`
+    const declared = declaredMimeOf(publicPath)
+
+    let status: 'ready' | 'quarantined' | 'unsupported'
+    let detectedMime: string
+    if (detected === null) {
+      status = 'unsupported'
+      detectedMime = UNKNOWN_DETECTED_MIME
+      issues.push(
+        `媒体未知签名（${row.sourceType}/${row.sourceId}，${publicPath}）：头部字节不匹配签名表——status=unsupported，不猜格式`,
+      )
+    } else if (detected.svg) {
+      status = 'quarantined'
+      detectedMime = detected.type
+      counts.mediaQuarantined++
+      issues.push(
+        `SVG 检出已隔离（${row.sourceType}/${row.sourceId}，${publicPath}）：文本 <svg 特征命中——status=quarantined，不进产物语义（spec §11.1）`,
+      )
+    } else {
+      status = 'ready'
+      detectedMime = detected.type
+      if (declared !== null && declared !== detected.type) {
+        issues.push(
+          `媒体声明与检测不符（${row.sourceType}/${row.sourceId}，${publicPath}）：扩展名推断 ${declared} ≠ 魔数检测 ${detected.type}——按检测登记（ready），数据质量问题已报告`,
+        )
+      }
+    }
+
+    try {
+      const inserted = await db
+        .insert(mediaAssets)
+        .values({
+          siteId,
+          sourceConnectionId: identity.connectionId,
+          sourceId: row.sourceId,
+          sourceUrl: fileUrl,
+          sourcePath: relPath ?? pathname.replace(/^\//, ''),
+          storagePath,
+          publicPath,
+          sha256: entry.sha256,
+          detectedMimeType: detectedMime,
+          declaredMimeType: declared,
+          byteSize: entry.bytes,
+          status,
+        })
+        // 唯一 (site, source_url) 定向幂等：命中零增行；public_path 唯一键冲突**不放行**
+        // （同公开路径双源是真实冲突，须显式报告而非静默吞掉）
+        .onConflictDoNothing({ target: [mediaAssets.siteId, mediaAssets.sourceUrl] })
+        .returning({ id: mediaAssets.id })
+      if (inserted.length > 0) {
+        counts.mediaRegistered++
+        if (fromSidecar) counts.mediaFromSidecar++
+      }
+    } catch (err) {
+      if (isUniqueViolation(err)) {
+        issues.push(
+          `媒体 public_path 冲突（${row.sourceType}/${row.sourceId}，${publicPath}）：同站公开路径已被他行占用——未登记（不静默覆盖）`,
+        )
+        return
+      }
+      throw err
+    }
+  }
+
+  /**
+   * 内容行引用提取（usage 两阶段的第一阶段）：refs + bodyHtml img src → 悬置引用。
+   * featured/gallery 取 id 锚（WP meta 的 _thumbnail_id/_product_image_gallery）；未知
+   * role 以 refs 原始下标落位（field_path 稳定优先于语义命名）。
+   */
+  function extractRefs(row: UniversalRow): Array<Omit<PendingRef, 'revisionId'>> {
+    const out: Array<Omit<PendingRef, 'revisionId'>> = []
+    let galleryOrdinal = 0
+    row.attachmentRefs.forEach((ref, i) => {
+      let fieldPath: string
+      if (ref.role === 'featured') fieldPath = 'attachmentRefs.featured'
+      else if (ref.role === 'gallery') fieldPath = `attachmentRefs.gallery[${galleryOrdinal++}]`
+      else fieldPath = `attachmentRefs.raw[${i}]`
+      if (ref.sourceId !== undefined && ref.sourceId !== '') {
+        out.push({ fieldPath, kind: 'id', value: ref.sourceId, uploadsPublicPath: null, rawLocation: row.rawLocation })
+      } else if (ref.sourceUrl !== undefined && ref.sourceUrl !== '') {
+        out.push({
+          fieldPath,
+          kind: 'url',
+          value: ref.sourceUrl,
+          uploadsPublicPath: uploadsPublicPathOf(ref.sourceUrl),
+          rawLocation: row.rawLocation,
+        })
+      }
+      // 双锚皆空：adapter 层已按非法 refs 记 issue（事件 issues 面），此处不重复报
+    })
+    const body = row.bodyHtml ?? ''
+    if (body !== '') {
+      const imgRe = /<img\b[^>]*?\bsrc\s*=\s*["']([^"']+)["']/gi
+      let m: RegExpExecArray | null
+      let imgOrdinal = 0
+      while ((m = imgRe.exec(body)) !== null) {
+        const src = m[1]!
+        if (!src.startsWith('data:')) {
+          // data: URI 是内联数据非资产引用；其余（绝对/相对/协议相对）统一进解析
+          out.push({
+            fieldPath: `data.bodyHtml.img[${imgOrdinal++}]`,
+            kind: 'url',
+            value: src,
+            uploadsPublicPath: uploadsPublicPathOf(src),
+            rawLocation: row.rawLocation,
+          })
+        }
+      }
+    }
+    return out
+  }
+
+  /** URL → 本站 uploads 公开路径形态（无标记 → null：外域热链不属悬空裁决域） */
+  function uploadsPublicPathOf(url: string): string | null {
+    const idx = url.indexOf(UPLOADS_MARKER)
+    return idx === -1 ? null : url.slice(idx)
+  }
+
+  /** 修订锚点：source_records 六列身份 → entry_locales.current_revision_id（content-writer 接缝） */
+  async function resolveRevisionAnchor(row: UniversalRow): Promise<string | null> {
+    const recs = await db
+      .select({ contentEntryId: sourceRecords.contentEntryId })
+      .from(sourceRecords)
+      .where(
+        and(
+          eq(sourceRecords.siteId, siteId),
+          eq(sourceRecords.sourceConnectionId, identity.connectionId),
+          eq(sourceRecords.sourceType, row.sourceType),
+          eq(sourceRecords.sourceId, row.sourceId),
+          eq(sourceRecords.locale, row.locale),
+          eq(sourceRecords.sourceSnapshotId, identity.snapshotId),
+        ),
+      )
+      .limit(1)
+    const entryId = recs[0]?.contentEntryId
+    if (entryId === undefined || entryId === null) return null
+    const els = await db
+      .select({ currentRevisionId: entryLocales.currentRevisionId })
+      .from(entryLocales)
+      .where(
+        and(
+          eq(entryLocales.siteId, siteId),
+          eq(entryLocales.contentEntryId, entryId),
+          eq(entryLocales.locale, row.locale),
+        ),
+      )
+      .limit(1)
+    return els[0]?.currentRevisionId ?? null
+  }
+
+  /** 内容行（非 attachment）：locale 口径对齐 content-writer（拒收行不双报），提取引用挂锚 */
+  async function handleContentRowRefs(row: UniversalRow): Promise<void> {
+    if (row.locale !== identity.defaultLocale) return
+    const refs = extractRefs(row)
+    if (refs.length === 0) return
+    const anchor = await resolveRevisionAnchor(row)
+    if (anchor === null) {
+      issues.push(
+        `媒体引用行无修订锚点（${row.sourceType}/${row.sourceId}/${row.locale}，refs=${refs.length}）：source_records/entry_locales 无 current 指针（content-writer 未处理该行？）——usage 跳过`,
+      )
+      return
+    }
+    for (const ref of refs) pendingRefs.push({ ...ref, revisionId: anchor })
+  }
+
+  /** 悬置引用 → 资产反查（id → source_id；URL → source_url 精确 → public_path 形态） */
+  async function lookupAsset(
+    ref: PendingRef,
+  ): Promise<{ id: string; sourceUrl: string; status: string } | null> {
+    const columns = {
+      id: mediaAssets.id,
+      sourceUrl: mediaAssets.sourceUrl,
+      status: mediaAssets.status,
+    }
+    if (ref.kind === 'id') {
+      const rows = await db
+        .select(columns)
+        .from(mediaAssets)
+        .where(
+          and(
+            eq(mediaAssets.siteId, siteId),
+            eq(mediaAssets.sourceConnectionId, identity.connectionId),
+            eq(mediaAssets.sourceId, ref.value),
+          ),
+        )
+        .limit(1)
+      return rows[0] ?? null
+    }
+    const byUrl = await db
+      .select(columns)
+      .from(mediaAssets)
+      .where(and(eq(mediaAssets.siteId, siteId), eq(mediaAssets.sourceUrl, ref.value)))
+      .limit(1)
+    if (byUrl.length > 0) return byUrl[0]!
+    if (ref.uploadsPublicPath !== null) {
+      const byPath = await db
+        .select(columns)
+        .from(mediaAssets)
+        .where(and(eq(mediaAssets.siteId, siteId), eq(mediaAssets.publicPath, ref.uploadsPublicPath)))
+        .limit(1)
+      return byPath[0] ?? null
+    }
+    return null
+  }
+
+  return {
+    async handleEvent(ev) {
+      if (ev.kind !== 'row') return // term/unsupported/error：无媒体信息，本写入器不消费
+      if (ev.row.sourceType === 'wp:attachment') {
+        await handleAttachmentRow(ev.row)
+        return
+      }
+      await handleContentRowRefs(ev.row)
+    },
+
+    /** 流终解析：悬置引用逐条对资产反查——命中且 ready → usage；无资产且属本域 → 悬空报告 */
+    async finish() {
+      for (const ref of pendingRefs) {
+        const asset = await lookupAsset(ref)
+        if (asset === null) {
+          if (ref.kind === 'id' || ref.uploadsPublicPath !== null) {
+            counts.mediaDanglingRefs++
+            const target = ref.kind === 'id' ? `attachment ${ref.value}` : ref.value
+            issues.push(
+              `媒体引用悬空（${ref.rawLocation}，${ref.fieldPath} → ${target}）：无对应 attachment/清单记录——引用保留于报告，不静默丢弃`,
+            )
+          }
+          // 外域 URL（无 uploads 标记）：站外热链不属本域对拍范围，静默越过（JSDoc 模块头）
+          continue
+        }
+        if (asset.status !== 'ready') {
+          // quarantined/unsupported 资产：不进产物语义——引用关系不建 usage（资产侧
+          // issue 已在登记时报告；引用命中非悬空）
+          continue
+        }
+        await db
+          .insert(mediaUsages)
+          .values({
+            siteId,
+            mediaAssetId: asset.id,
+            contentRevisionId: ref.revisionId,
+            fieldPath: ref.fieldPath,
+            // id 引用无引文 URL——落资产登记的 source_url（审计可追）；URL 引用落引文原串
+            sourceUrl: ref.kind === 'url' ? ref.value : asset.sourceUrl,
+          })
+          .onConflictDoNothing()
+      }
+      return { counts, issues }
+    },
+  }
+}
+
+/**
+ * 从管道上下文构造媒体写入器句柄（Task 6 编排注入位）：manifest/sidecar 从快照 metadata
+ * 与存储副本解析（media_bundle.metadata.manifest / manual_url_list 副本 JSON）。
+ * bundle 缺席 = 空清单（三级全走 missing——诚实报告而非跳过）；侧车缺席 = 空映射。
+ */
+export async function createMediaEventHandlerFromContext(
+  ctx: Pick<
+    ImportJobContext,
+    | 'db'
+    | 'siteId'
+    | 'connectionId'
+    | 'defaultLocale'
+    | 'wxrSnapshot'
+    | 'mediaBundleSnapshot'
+    | 'manualUrlListSnapshot'
+  >,
+): Promise<MediaEventHandler> {
+  const bundle = ctx.mediaBundleSnapshot
+  const manifestRaw = bundle?.metadata.manifest
+  const manifest: MediaManifestEntry[] = Array.isArray(manifestRaw) ? (manifestRaw as MediaManifestEntry[]) : []
+  let storageRoot = ''
+  if (manifest.length > 0) {
+    const raw = bundle?.metadata.storageRoot
+    if (typeof raw !== 'string' || raw === '') {
+      throw new Error('media_bundle 快照 metadata 缺 storageRoot（登记期事实缺席——无法读取存储副本做魔数校验）')
+    }
+    storageRoot = raw
+  }
+  return createMediaEventHandler({
+    db: ctx.db,
+    siteId: ctx.siteId,
+    connectionId: ctx.connectionId,
+    snapshotId: ctx.wxrSnapshot.id,
+    defaultLocale: ctx.defaultLocale,
+    storageRoot,
+    mediaBundleStoragePath: bundle?.storagePath ?? '',
+    mediaManifest: manifest,
+    sidecarMap: await loadSidecarMap(ctx),
+  })
+}
+
+/** 侧车加载：manual_url_list 副本 JSON → Map（source_url → file_path）。形状在登记期已校验 */
+async function loadSidecarMap(
+  ctx: Pick<ImportJobContext, 'manualUrlListSnapshot'>,
+): Promise<Map<string, string>> {
+  const snap = ctx.manualUrlListSnapshot
+  if (snap === null) return new Map()
+  const root = snap.metadata.storageRoot
+  if (typeof root !== 'string' || root === '') {
+    throw new Error('manual_url_list 快照 metadata 缺 storageRoot（登记期事实缺席——无法读取侧车副本）')
+  }
+  let parsed: unknown
+  try {
+    parsed = JSON.parse(await readFile(join(root, snap.storagePath), 'utf8'))
+  } catch (err) {
+    throw new Error(
+      `侧车快照副本解析失败（${snap.storagePath}）：${err instanceof Error ? err.message : String(err)}`,
+    )
+  }
+  if (!Array.isArray(parsed)) {
+    throw new Error(`侧车快照副本形状异常（${snap.storagePath}）：顶层不是数组`)
+  }
+  const map = new Map<string, string>()
+  for (const item of parsed) {
+    const entry = item as Partial<SidecarEntry>
+    if (
+      typeof entry.sourceUrl !== 'string' ||
+      entry.sourceUrl === '' ||
+      typeof entry.filePath !== 'string' ||
+      entry.filePath === ''
+    ) {
+      throw new Error(`侧车快照副本形状异常（${snap.storagePath}）：存在非 {sourceUrl, filePath} 条目`)
+    }
+    map.set(entry.sourceUrl, entry.filePath)
+  }
+  return map
+}
+
+/**
+ * 媒体登记管道写入器（{@link createMediaEventHandler} 的自消费整流便捷壳）：单遍拉完
+ * ctx.events、逐事件分派、流终返回 outcome。与 content-writer 组合消费同流时须用
+ * {@link createMediaEventHandlerFromContext} 句柄形状（单消费方拉流分派，无 tee）。
+ */
+export function createMediaWriter(): ImportPipeline {
+  return async (ctx) => {
+    const handler = await createMediaEventHandlerFromContext(ctx)
+    for await (const ev of ctx.events) {
+      await handler.handleEvent(ev)
+    }
+    return handler.finish()
+  }
+}
diff --git a/tests/integration/import-media.test.ts b/tests/integration/import-media.test.ts
new file mode 100644
index 0000000..0d582f8
--- /dev/null
+++ b/tests/integration/import-media.test.ts
@@ -0,0 +1,643 @@
+// Plan 3 Task 4 失败测试（先于实现）：媒体登记写入器（三级对拍 + 魔数/隔离 + 悬空引用）。
+// 覆盖契约（对照计划 Task 4 与前置事实 7/11、spec §5.5/§11.1/§7.6）：
+// 1. mini 全链三级对拍：①uploads 相对路径清单精确命中（fixture-a）②侧车 source_url→
+//    file_path→清单实体（fixture-b——清单故意 miss 该路径逼走侧车级）③missing 不插行
+//    （media_assets sha256 NOT NULL——另测无 bundle 场景）；命中行 public_path=旧 uploads
+//    路径、sha256/byte_size/storage_path 从清单取；media_usages 三行（featured/gallery/
+//    正文 img——从 row.attachmentRefs + bodyHtml 提取，revision 锚点经 source_records→
+//    entry_locales.current_revision_id）；
+// 2. 魔数（集成面）：命中文件读实际字节——PNG 过；扩展名声明 MIME 与检测不符 → issue；
+//    SVG 检出（.png 伪装与显式 .svg 双例）→ status='quarantined' + issue（不进产物语义）；
+//    未知签名 → status='unsupported' + issue；quarantined 资产的引用不建 usage；
+// 3. 悬空引用（mini 203/204）：gallery 引用无对应 attachment/清单记录 → mediaDanglingRefs
+//    计数 + issue（不静默丢弃，§7.6）；
+// 4. 多候选：侧车 file_path 按文件名命中两条清单（同 filename 不同目录）→ 歧义不自动选、
+//    不插行、全部候选列 issue（spec §5.5）；
+// 5. 幂等：重跑 media_assets/media_usages 零增行（mediaRegistered=0；悬空引用按流如实再现）；
+// 6. counts 键（media 前缀全局命名空间）：mediaRegistered/mediaFromSidecar/mediaMissing/
+//    mediaQuarantined/mediaDanglingRefs/mediaAmbiguous——与 content-writer 键并存无碰撞。
+// 双事件消费设计（实现侧 JSDoc 同步）：attachment 行建资产、内容行提取引用；WXR 文档序
+// items 先于 attachment——引用先行资产后到，usage 解析在 finish() 统一对 media_assets
+// 反查（顺序无关）。运行于 globalSetup 临时库（ADR-015）；报告根 os.tmpdir，finally 清理。
+import { createHash } from 'node:crypto'
+import { createReadStream } from 'node:fs'
+import { mkdir, mkdtemp, readFile, rm, writeFile } from 'node:fs/promises'
+import { tmpdir } from 'node:os'
+import { join, resolve } from 'node:path'
+import { fileURLToPath } from 'node:url'
+import { describe, expect, it } from 'vitest'
+import { runWxrAdapter, type AdapterEvent } from '@sf/adapter-wp'
+import type { UniversalRow } from '@sf/contracts'
+import type { MediaManifestEntry } from '../../apps/api/src/modules/source/register.service.js'
+import type { ImportPipeline } from '../../apps/api/src/modules/import/import.service.js'
+import { createImportJob } from '../../apps/api/src/modules/import/import.service.js'
+import { createContentEventHandler } from '../../apps/api/src/modules/import/content-writer.js'
+import { createMediaEventHandler, createMediaEventHandlerFromContext } from '../../apps/api/src/modules/import/media-writer.js'
+import { registerSource } from '../../apps/api/src/modules/source/register.service.js'
+import { rand, withImmutableCleanup, withTestDb, type Db } from './helpers.js'
+
+const FIXTURES_ROOT = fileURLToPath(new URL('../fixtures', import.meta.url))
+const WXR_FIXTURE = resolve(FIXTURES_ROOT, 'wxr', 'mini.xml')
+
+const PNG_SIG = Buffer.from([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a])
+const JPEG_SIG = Buffer.from([0xff, 0xd8, 0xff, 0xe0])
+
+/** 伪 PNG/JPEG 体（签名 + 唯一种子载荷——魔数只看头，内容唯一避免跨测试清单撞哈希） */
+function pngBytes(seed: string): Buffer {
+  return Buffer.concat([PNG_SIG, Buffer.from(`png-body:${seed}`, 'latin1')])
+}
+function jpegBytes(seed: string): Buffer {
+  return Buffer.concat([JPEG_SIG, Buffer.from(`jpeg-body:${seed}`, 'latin1')])
+}
+function sha256of(buf: Buffer): string {
+  return createHash('sha256').update(buf).digest('hex')
+}
+function manifestEntryOf(path: string, buf: Buffer): MediaManifestEntry {
+  return { path, bytes: buf.length, sha256: sha256of(buf) }
+}
+
+/** 手构事件流的 AsyncIterable 包装（每次调用产新迭代器——重跑各流独立） */
+async function* eventsOf(events: AdapterEvent[]): AsyncGenerator<AdapterEvent, void, void> {
+  yield* events
+}
+
+/** db.execute 行对象规整（postgres.js 原型差异会让 toStrictEqual 报"无视觉差异"失败） */
+function plain(rows: any[]): any[] {
+  return rows.map((r) => ({ ...r }))
+}
+
+/** 事件流收集（mini 全链一饮而尽） */
+async function collectMiniEvents(): Promise<AdapterEvent[]> {
+  const out: AdapterEvent[] = []
+  for await (const ev of runWxrAdapter(createReadStream(WXR_FIXTURE), { mapConfig: { locale: 'en' } })) {
+    out.push(ev)
+  }
+  return out
+}
+
+/** 手构 attachment row 事件（knownFields.attachmentUrl 是三级对拍的输入锚点） */
+function attachmentEvent(spec: { sourceId: string; url: string }): AdapterEvent {
+  const row: UniversalRow = {
+    sourceType: 'wp:attachment',
+    sourceId: spec.sourceId,
+    status: 'inherit',
+    locale: 'en',
+    taxonomyRefs: [],
+    attachmentRefs: [],
+    seo: {},
+    knownFields: { attachmentUrl: spec.url },
+    extras: {},
+    rawLocation: `wxr://item/${spec.sourceId}`,
+  }
+  return { kind: 'row', row, issues: [], sanitizeDiffs: [], changed: false }
+}
+
+/** Task 6 编排雏形：单消费方拉流，逐事件先 content 后 media 分派（content 先行是
+ * media-writer usage 锚点的顺序契约——source_records/revision 须已落）；两态经 finish 取回 */
+function contentPlusMediaPipeline(): ImportPipeline {
+  return contentPlusDispatch(async (ctx) => createMediaEventHandlerFromContext(ctx))
+}
+
+/** content + 直传 media 组合（手构场景：不经 register 的 media_bundle）——分派形状同上 */
+function contentPlusDirectMediaPipeline(identity: {
+  storageRoot: string
+  manifest: MediaManifestEntry[]
+  sidecar: Array<[string, string]>
+}): ImportPipeline {
+  return contentPlusDispatch(async (ctx) =>
+    createMediaEventHandler({
+      db: ctx.db,
+      siteId: ctx.siteId,
+      connectionId: ctx.connectionId,
+      snapshotId: ctx.wxrSnapshot.id,
+      defaultLocale: ctx.defaultLocale,
+      storageRoot: identity.storageRoot,
+      mediaBundleStoragePath: '',
+      mediaManifest: identity.manifest,
+      sidecarMap: new Map(identity.sidecar),
+    }),
+  )
+}
+
+/** 分派骨架：先 content 后 media（顺序契约），counts 展开合并（media 前缀无碰撞键） */
+function contentPlusDispatch(
+  mediaOf: (ctx: Parameters<ImportPipeline>[0]) => Promise<ReturnType<typeof createMediaEventHandler>>,
+): ImportPipeline {
+  return async (ctx) => {
+    const content = createContentEventHandler({
+      db: ctx.db,
+      siteId: ctx.siteId,
+      connectionId: ctx.connectionId,
+      snapshotId: ctx.wxrSnapshot.id,
+      jobId: ctx.jobId,
+      defaultLocale: ctx.defaultLocale,
+      versions: ctx.versions,
+    })
+    const media = await mediaOf(ctx)
+    for await (const ev of ctx.events) {
+      await content.handleEvent(ev)
+      await media.handleEvent(ev)
+    }
+    const c = content.finish()
+    const m = await media.finish()
+    return { counts: { ...c.counts, ...m.counts }, issues: [...c.issues, ...m.issues] }
+  }
+}
+
+/** 清单/侧车直传管道（手构场景：无内容行——纯 attachment 事件，media 单写入器） */
+function directMediaPipeline(identity: {
+  storageRoot: string
+  manifest: MediaManifestEntry[]
+  sidecar: Array<[string, string]>
+}): ImportPipeline {
+  return async (ctx) => {
+    const handler = createMediaEventHandler({
+      db: ctx.db,
+      siteId: ctx.siteId,
+      connectionId: ctx.connectionId,
+      snapshotId: ctx.wxrSnapshot.id,
+      defaultLocale: ctx.defaultLocale,
+      storageRoot: identity.storageRoot,
+      mediaBundleStoragePath: '',
+      mediaManifest: identity.manifest,
+      sidecarMap: new Map(identity.sidecar),
+    })
+    for await (const ev of ctx.events) {
+      await handler.handleEvent(ev)
+    }
+    return handler.finish()
+  }
+}
+
+/** 本文件写入的行清场：FK 反序（media_usages → media_assets 先于内容域；两张不可变表走触发器开关） */
+async function cleanupSiteRows(db: Db, siteId: string): Promise<void> {
+  await db.execute(`delete from media_usages where site_id = '${siteId}'`)
+  await db.execute(`delete from media_assets where site_id = '${siteId}'`)
+  await db.execute(`delete from content_terms where site_id = '${siteId}'`)
+  await db.execute(`delete from terms where site_id = '${siteId}'`)
+  await db.execute(`delete from taxonomies where site_id = '${siteId}'`)
+  await db.execute(`delete from source_records where site_id = '${siteId}'`)
+  await db.execute(`delete from entry_locales where site_id = '${siteId}'`)
+  await withImmutableCleanup(db, 'content_revisions', 'content_revisions_immutable', `site_id = '${siteId}'`)
+  await db.execute(`delete from content_entries where site_id = '${siteId}'`)
+  await db.execute(`delete from import_jobs where site_id = '${siteId}'`)
+  await withImmutableCleanup(db, 'source_snapshots', 'source_snapshots_immutable', `site_id = '${siteId}'`)
+  await db.execute(`delete from source_connections where site_id = '${siteId}'`)
+  await db.execute(`delete from sites where id = '${siteId}'`)
+}
+
+/** mini 登记（各用例公共前置）：register → {siteId, connectionId, wxr 快照} */
+async function registerMini(db: Db, work: string) {
+  const key = `sfmedia-${rand()}`
+  const reg = await registerSource(db, {
+    siteKey: key,
+    wxrPath: WXR_FIXTURE,
+    allowedRoots: [FIXTURES_ROOT, work],
+    storageRoot: join(work, 'storage'),
+    reportRoot: join(work, 'artifacts', 'register'),
+  })
+  return { key, reg, wxr: reg.snapshots.find((s) => s.kind === 'wxr')! }
+}
+
+/** 导入报告读取 */
+async function readReport(work: string, jobId: string): Promise<{
+  status: string
+  counts: Record<string, number>
+  issues: string[]
+}> {
+  return JSON.parse(await readFile(join(work, 'artifacts', 'import', jobId, 'reports', 'import.json'), 'utf8'))
+}
+
+describe('媒体登记写入器（三级对拍 + 魔数/隔离）', () => {
+  it('mini 全链三级对拍：①清单精确命中 ②侧车命中 → media_assets（public_path=旧 uploads 路径、sha256/storage_path 从清单取）+ usage 三行 + 悬空 203/204 + counts 全键', async () => {
+    await withTestDb(async (db) => {
+      const work = await mkdtemp(join(tmpdir(), 'sf-media-full-'))
+      try {
+        const events = await collectMiniEvents()
+        // 媒体目录：fixture-a 按旧 uploads 层级放（级①精确命中）；fixture-b 用扁平
+        // GUID 名（清单故意不含 '2026/10/fixture-b.png'——逼走级②侧车）
+        const mediaDir = join(work, 'media')
+        await mkdir(join(mediaDir, '2026/10'), { recursive: true })
+        const pngA = pngBytes(`fixture-a-${rand()}`)
+        await writeFile(join(mediaDir, '2026/10/fixture-a.png'), pngA)
+        const guidB = `b-${rand()}.png`
+        const pngB = pngBytes(`fixture-b-${rand()}`)
+        await writeFile(join(mediaDir, guidB), pngB)
+        const sidecarPath = join(work, 'media-map.json')
+        await writeFile(
+          sidecarPath,
+          JSON.stringify([
+            { sourceUrl: 'https://fixture.example/wp-content/uploads/2026/10/fixture-b.png', filePath: guidB },
+          ]),
+        )
+        const key = `sfmedia-${rand()}`
+        const reg = await registerSource(db, {
+          siteKey: key,
+          wxrPath: WXR_FIXTURE,
+          mediaDir,
+          mediaMapPath: sidecarPath,
+          allowedRoots: [FIXTURES_ROOT, work],
+          storageRoot: join(work, 'storage'),
+          reportRoot: join(work, 'artifacts', 'register'),
+        })
+        const wxr = reg.snapshots.find((s) => s.kind === 'wxr')!
+        const bundle = reg.snapshots.find((s) => s.kind === 'media_bundle')!
+        try {
+          const result = await createImportJob(db, {
+            siteKey: key,
+            snapshotId: wxr.id,
+            reportRoot: join(work, 'artifacts', 'import'),
+            events: eventsOf(events),
+            pipeline: contentPlusMediaPipeline(),
+          })
+          // 悬空 203/204 → issues 非空 → succeeded_with_issues（§7.6 禁静默）
+          expect(result.status).toBe('succeeded_with_issues')
+
+          // media_assets 两行：fixture-a 级① / fixture-b 级②；身份列逐项对拍
+          const assets = (await db.execute(`select source_id, source_url, source_path, storage_path,
+            public_path, sha256, detected_mime, declared_mime, byte_size::int as byte_size, status
+            from media_assets where site_id = '${reg.siteId}' order by source_id`)) as any
+          expect(assets).toHaveLength(2)
+          const a = assets.find((x: any) => x.source_id === '201')!
+          expect(a.source_url).toBe('https://fixture.example/wp-content/uploads/2026/10/fixture-a.png')
+          expect(a.source_path).toBe('2026/10/fixture-a.png')
+          expect(a.public_path).toBe('/wp-content/uploads/2026/10/fixture-a.png')
+          expect(a.sha256).toBe(sha256of(pngA))
+          expect(a.byte_size).toBe(pngA.length)
+          expect(a.detected_mime).toBe('image/png')
+          expect(a.declared_mime).toBe('image/png')
+          expect(a.status).toBe('ready')
+          // storage_path = 清单实体定位（bundle 副本相对路径 + 清单 path——绝对副本 = storageRoot + 此列）
+          expect(a.storage_path).toBe(`${bundle.storagePath}/2026/10/fixture-a.png`)
+          const b = assets.find((x: any) => x.source_id === '202')!
+          expect(b.source_url).toBe('https://fixture.example/wp-content/uploads/2026/10/fixture-b.png')
+          expect(b.public_path).toBe('/wp-content/uploads/2026/10/fixture-b.png')
+          expect(b.sha256).toBe(sha256of(pngB))
+          expect(b.storage_path).toBe(`${bundle.storagePath}/${guidB}`)
+          expect(b.status).toBe('ready')
+
+          // media_usages 三行：101 featured（id 反查）+ 101 正文 img（URL 反查）+ 111 featured
+          const usages = (await db.execute(`select a.source_id as asset, u.field_path, u.source_url
+            from media_usages u join media_assets a on a.id = u.media_asset_id
+            where u.site_id = '${reg.siteId}' order by a.source_id, u.field_path`)) as any
+          expect(plain(usages)).toStrictEqual([
+            {
+              asset: '201',
+              field_path: 'attachmentRefs.featured',
+              source_url: 'https://fixture.example/wp-content/uploads/2026/10/fixture-a.png',
+            },
+            {
+              asset: '201',
+              field_path: 'data.bodyHtml.img[0]',
+              source_url: 'https://fixture.example/wp-content/uploads/2026/10/fixture-a.png',
+            },
+            {
+              asset: '202',
+              field_path: 'attachmentRefs.featured',
+              source_url: 'https://fixture.example/wp-content/uploads/2026/10/fixture-b.png',
+            },
+          ])
+          // usage 挂在 content-writer 落的当前修订上（current_revision_id——导入不发布，published NULL）
+          const revCheck = (await db.execute(`select count(*)::int as n from media_usages u
+            join content_revisions cr on cr.id = u.content_revision_id
+            join entry_locales el on el.current_revision_id = cr.id and el.published_revision_id is null
+            where u.site_id = '${reg.siteId}'`)) as any
+          expect(revCheck[0].n).toBe(3)
+
+          // counts 全键：content 四键 + media 六键（全局命名空间并存无碰撞）
+          const report = await readReport(work, result.jobId)
+          expect(report.status).toBe('succeeded_with_issues')
+          expect(report.counts).toStrictEqual({
+            contentEntries: 3,
+            revisionsCreated: 3,
+            sourceRecords: 5,
+            skipped: 0,
+            mediaRegistered: 2,
+            mediaFromSidecar: 1,
+            mediaMissing: 0,
+            mediaQuarantined: 0,
+            // mini 全部引用 5 条：featured 201/202 + 正文 img 各解析命中，gallery 203/204 悬空
+            mediaDanglingRefs: 2,
+            mediaAmbiguous: 0,
+          })
+          // 悬空引用显式报告：203/204 各一条（gallery 引用无 attachment/清单记录）
+          const dangling = report.issues.filter((i) => i.includes('悬空'))
+          expect(dangling).toHaveLength(2)
+          expect(dangling.some((i) => i.includes('203'))).toBe(true)
+          expect(dangling.some((i) => i.includes('204'))).toBe(true)
+        } finally {
+          await cleanupSiteRows(db, reg.siteId)
+        }
+      } finally {
+        await rm(work, { recursive: true, force: true })
+      }
+    })
+  })
+
+  it('级③ missing（无 media_bundle 登记）：不插 media_assets 行（sha256 NOT NULL）+ mediaMissing + 全部引用悬空如实报告', async () => {
+    await withTestDb(async (db) => {
+      const work = await mkdtemp(join(tmpdir(), 'sf-media-missing-'))
+      try {
+        const events = await collectMiniEvents()
+        const { key, reg, wxr } = await registerMini(db, work)
+        try {
+          const result = await createImportJob(db, {
+            siteKey: key,
+            snapshotId: wxr.id,
+            reportRoot: join(work, 'artifacts', 'import'),
+            events: eventsOf(events),
+            pipeline: contentPlusMediaPipeline(),
+          })
+          expect(result.status).toBe('succeeded_with_issues')
+
+          // 零资产行：清单缺席 → 级①②全 miss → 级③ missing（不插行——sha256 无来源）
+          expect(((await db.execute(`select count(*)::int as n from media_assets
+            where site_id = '${reg.siteId}'`)) as any)[0].n).toBe(0)
+          expect(((await db.execute(`select count(*)::int as n from media_usages
+            where site_id = '${reg.siteId}'`)) as any)[0].n).toBe(0)
+
+          const report = await readReport(work, result.jobId)
+          expect(report.counts).toStrictEqual({
+            contentEntries: 3,
+            revisionsCreated: 3,
+            sourceRecords: 5,
+            skipped: 0,
+            mediaRegistered: 0,
+            mediaFromSidecar: 0,
+            mediaMissing: 2,
+            mediaQuarantined: 0,
+            // 引用全悬空：101 featured+img、111 featured+gallery×2 = 5
+            mediaDanglingRefs: 5,
+            mediaAmbiguous: 0,
+          })
+          const missing = report.issues.filter((i) => i.includes('missing'))
+          expect(missing).toHaveLength(2)
+          expect(missing.some((i) => i.includes('fixture-a.png'))).toBe(true)
+          expect(missing.some((i) => i.includes('fixture-b.png'))).toBe(true)
+          expect(report.issues.filter((i) => i.includes('悬空'))).toHaveLength(5)
+        } finally {
+          await cleanupSiteRows(db, reg.siteId)
+        }
+      } finally {
+        await rm(work, { recursive: true, force: true })
+      }
+    })
+  })
+
+  it('魔数校验：命中文件读实际字节——PNG 过；声明（扩展名推断）与检测不符 → issue；SVG 检出（伪装 .png 与显式 .svg）→ quarantined 不进产物；未知签名 → unsupported', async () => {
+    await withTestDb(async (db) => {
+      const work = await mkdtemp(join(tmpdir(), 'sf-media-magic-'))
+      try {
+        // 五个手构文件（放旧 uploads 层级——级① 相对路径精确命中）：正例 PNG / 扩展名谎言
+        // （.png 实为 JPEG）/ SVG 伪装 .png / 显式 .svg / 无签名 .txt
+        const files: Array<{ name: string; buf: Buffer }> = [
+          { name: 'ok.png', buf: pngBytes(`ok-${rand()}`) },
+          { name: 'lie.png', buf: jpegBytes(`lie-${rand()}`) },
+          { name: 'evil.png', buf: Buffer.from(`<?xml version="1.0"?><SVG onload="alert(1)" xmlns="…">`) },
+          { name: 'draw.svg', buf: Buffer.from('<svg width="10" height="10"></svg>') },
+          { name: 'note.txt', buf: Buffer.from('plain text note, no signature at all') },
+        ]
+        await mkdir(join(work, '2026/10'), { recursive: true })
+        for (const f of files) {
+          await writeFile(join(work, '2026/10', f.name), f.buf)
+        }
+        const manifest = files.map((f) => manifestEntryOf(`2026/10/${f.name}`, f.buf))
+        const events: AdapterEvent[] = files.map((f, i) =>
+          attachmentEvent({ sourceId: String(901 + i), url: `https://fixture.example/wp-content/uploads/2026/10/${f.name}` }),
+        )
+        const { key, reg, wxr } = await registerMini(db, work)
+        try {
+          const result = await createImportJob(db, {
+            siteKey: key,
+            snapshotId: wxr.id,
+            reportRoot: join(work, 'artifacts', 'import'),
+            events: eventsOf(events),
+            pipeline: directMediaPipeline({ storageRoot: work, manifest, sidecar: [] }),
+          })
+          expect(result.status).toBe('succeeded_with_issues')
+
+          const assets = (await db.execute(`select source_id, detected_mime, declared_mime, status
+            from media_assets where site_id = '${reg.siteId}' order by source_id`)) as any
+          expect(plain(assets)).toStrictEqual([
+            { source_id: '901', detected_mime: 'image/png', declared_mime: 'image/png', status: 'ready' },
+            { source_id: '902', detected_mime: 'image/jpeg', declared_mime: 'image/png', status: 'ready' },
+            { source_id: '903', detected_mime: 'image/svg+xml', declared_mime: 'image/png', status: 'quarantined' },
+            { source_id: '904', detected_mime: 'image/svg+xml', declared_mime: 'image/svg+xml', status: 'quarantined' },
+            { source_id: '905', detected_mime: 'application/octet-stream', declared_mime: null, status: 'unsupported' },
+          ])
+
+          const report = await readReport(work, result.jobId)
+          expect(report.counts).toStrictEqual({
+            mediaRegistered: 5,
+            mediaFromSidecar: 0,
+            mediaMissing: 0,
+            mediaQuarantined: 2,
+            mediaDanglingRefs: 0,
+            mediaAmbiguous: 0,
+          })
+          // 三类问题各就位：MIME 不符 1（lie.png）+ SVG 隔离 2（evil/draw）+ 未知签名 1（note.txt）
+          const mismatch = report.issues.filter((i) => i.includes('不符'))
+          expect(mismatch).toHaveLength(1)
+          expect(mismatch[0]).toContain('lie.png')
+          const quarantined = report.issues.filter((i) => i.includes('隔离'))
+          expect(quarantined).toHaveLength(2)
+          expect(report.issues.filter((i) => i.includes('未知签名'))).toHaveLength(1)
+          // quarantined/unsupported 资产不进产物语义：即便有引用也不建 usage（本例无内容行，零 usage）
+          expect(((await db.execute(`select count(*)::int as n from media_usages
+            where site_id = '${reg.siteId}'`)) as any)[0].n).toBe(0)
+        } finally {
+          await cleanupSiteRows(db, reg.siteId)
+        }
+      } finally {
+        await rm(work, { recursive: true, force: true })
+      }
+    })
+  })
+
+  it('quarantined 资产的引用不建 usage（不进产物语义）：SVG 资产被正文引用 → 引用关系不落 media_usages', async () => {
+    await withTestDb(async (db) => {
+      const work = await mkdtemp(join(tmpdir(), 'sf-media-quar-'))
+      try {
+        const evil = Buffer.from('<svg onload="x()"><script>bad()</script></svg>')
+        await mkdir(join(work, '2026/10'), { recursive: true })
+        await writeFile(join(work, '2026/10/evil.png'), evil)
+        const manifest = [manifestEntryOf('2026/10/evil.png', evil)]
+        const attachment = attachmentEvent({
+          sourceId: '951',
+          url: 'https://fixture.example/wp-content/uploads/2026/10/evil.png',
+        })
+        const citing: AdapterEvent = {
+          kind: 'row',
+          row: {
+            sourceType: 'wp:post',
+            sourceId: '950',
+            status: 'publish',
+            locale: 'en',
+            taxonomyRefs: [],
+            attachmentRefs: [{ role: 'featured', sourceId: '951' }],
+            seo: {},
+            knownFields: {},
+            extras: {},
+            rawLocation: 'wxr://item/950',
+            bodyHtml: '<p><img src="https://fixture.example/wp-content/uploads/2026/10/evil.png"/></p>',
+          },
+          issues: [],
+          sanitizeDiffs: [],
+          changed: false,
+        }
+        const { key, reg, wxr } = await registerMini(db, work)
+        try {
+          // 引用行先到、attachment 后到（WXR 文档序）——finish 统一解析
+          const result = await createImportJob(db, {
+            siteKey: key,
+            snapshotId: wxr.id,
+            reportRoot: join(work, 'artifacts', 'import'),
+            events: eventsOf([citing, attachment]),
+            pipeline: contentPlusDirectMediaPipeline({ storageRoot: work, manifest, sidecar: [] }),
+          })
+          expect(result.status).toBe('succeeded_with_issues')
+          // 资产行存在（登记可审计）但 quarantined；引用解析命中资产（非悬空）但不建 usage
+          const assets = (await db.execute(`select source_id, status from media_assets
+            where site_id = '${reg.siteId}'`)) as any
+          expect(plain(assets)).toStrictEqual([{ source_id: '951', status: 'quarantined' }])
+          expect(((await db.execute(`select count(*)::int as n from media_usages
+            where site_id = '${reg.siteId}'`)) as any)[0].n).toBe(0)
+          const report = await readReport(work, result.jobId)
+          expect(report.counts.mediaRegistered).toBe(1)
+          expect(report.counts.mediaQuarantined).toBe(1)
+          expect(report.counts.mediaDanglingRefs).toBe(0)
+          expect(report.issues.filter((i) => i.includes('隔离'))).toHaveLength(1)
+        } finally {
+          await cleanupSiteRows(db, reg.siteId)
+        }
+      } finally {
+        await rm(work, { recursive: true, force: true })
+      }
+    })
+  })
+
+  it('多候选（同 filename 不同目录两条清单）：歧义不自动选、不插行、全部候选列 issue', async () => {
+    await withTestDb(async (db) => {
+      const work = await mkdtemp(join(tmpdir(), 'sf-media-ambig-'))
+      try {
+        await mkdir(join(work, 'd1'), { recursive: true })
+        await mkdir(join(work, 'd2'), { recursive: true })
+        const f1 = pngBytes(`amb-1-${rand()}`)
+        const f2 = pngBytes(`amb-2-${rand()}`)
+        await writeFile(join(work, 'd1/thing.png'), f1)
+        await writeFile(join(work, 'd2/thing.png'), f2)
+        const manifest = [manifestEntryOf('d1/thing.png', f1), manifestEntryOf('d2/thing.png', f2)]
+        const url = 'https://fixture.example/wp-content/uploads/2024/01/thing.png'
+        const events: AdapterEvent[] = [attachmentEvent({ sourceId: '961', url })]
+        const { key, reg, wxr } = await registerMini(db, work)
+        try {
+          const result = await createImportJob(db, {
+            siteKey: key,
+            snapshotId: wxr.id,
+            reportRoot: join(work, 'artifacts', 'import'),
+            events: eventsOf(events),
+            // 侧车 file_path 只给文件名：精确路径 miss → 文件名候选 2 条 → 歧义
+            pipeline: directMediaPipeline({ storageRoot: work, manifest, sidecar: [[url, 'thing.png']] }),
+          })
+          expect(result.status).toBe('succeeded_with_issues')
+          expect(((await db.execute(`select count(*)::int as n from media_assets
+            where site_id = '${reg.siteId}'`)) as any)[0].n).toBe(0)
+          const report = await readReport(work, result.jobId)
+          expect(report.counts).toStrictEqual({
+            mediaRegistered: 0,
+            mediaFromSidecar: 0,
+            mediaMissing: 0,
+            mediaQuarantined: 0,
+            mediaDanglingRefs: 0,
+            mediaAmbiguous: 1,
+          })
+          const amb = report.issues.filter((i) => i.includes('歧义'))
+          expect(amb).toHaveLength(1)
+          expect(amb[0]).toContain('d1/thing.png')
+          expect(amb[0]).toContain('d2/thing.png')
+        } finally {
+          await cleanupSiteRows(db, reg.siteId)
+        }
+      } finally {
+        await rm(work, { recursive: true, force: true })
+      }
+    })
+  })
+
+  it('幂等重跑：media_assets/media_usages 零增行、mediaRegistered=0；悬空引用按流如实再现', async () => {
+    await withTestDb(async (db) => {
+      const work = await mkdtemp(join(tmpdir(), 'sf-media-idem-'))
+      try {
+        const events = await collectMiniEvents()
+        const mediaDir = join(work, 'media')
+        await mkdir(join(mediaDir, '2026/10'), { recursive: true })
+        const pngA = pngBytes(`idem-a-${rand()}`)
+        await writeFile(join(mediaDir, '2026/10/fixture-a.png'), pngA)
+        const guidB = `b-${rand()}.png`
+        await writeFile(join(mediaDir, guidB), pngBytes(`idem-b-${rand()}`))
+        const sidecarPath = join(work, 'media-map.json')
+        await writeFile(
+          sidecarPath,
+          JSON.stringify([
+            { sourceUrl: 'https://fixture.example/wp-content/uploads/2026/10/fixture-b.png', filePath: guidB },
+          ]),
+        )
+        const key = `sfmedia-${rand()}`
+        const reg = await registerSource(db, {
+          siteKey: key,
+          wxrPath: WXR_FIXTURE,
+          mediaDir,
+          mediaMapPath: sidecarPath,
+          allowedRoots: [FIXTURES_ROOT, work],
+          storageRoot: join(work, 'storage'),
+          reportRoot: join(work, 'artifacts', 'register'),
+        })
+        const wxr = reg.snapshots.find((s) => s.kind === 'wxr')!
+        try {
+          const base = {
+            siteKey: key,
+            snapshotId: wxr.id,
+            reportRoot: join(work, 'artifacts', 'import'),
+            pipeline: contentPlusMediaPipeline(),
+          }
+          const first = await createImportJob(db, { ...base, events: eventsOf(events) })
+          expect(first.status).toBe('succeeded_with_issues')
+
+          // 同流重跑（换幂等键新 job——钉写入器幂等，不测 job 复用）
+          const again = await createImportJob(db, {
+            ...base,
+            idempotencyKey: `sfmedia-rerun-${rand()}`,
+            events: eventsOf(events),
+          })
+          expect(again.reused).toBe(false)
+          expect(again.status).toBe('succeeded_with_issues')
+
+          const after = (await db.execute(`select
+            (select count(*)::int from media_assets where site_id = '${reg.siteId}') as assets,
+            (select count(*)::int from media_usages where site_id = '${reg.siteId}') as usages`)) as any
+          expect(plain(after)[0]).toStrictEqual({ assets: 2, usages: 3 })
+
+          // 行面零增；流面计数如实：registered/fromSidecar 归零，悬空引用再现（203/204）
+          const report = await readReport(work, again.jobId)
+          expect(report.counts).toStrictEqual({
+            contentEntries: 0,
+            revisionsCreated: 0,
+            sourceRecords: 0,
+            skipped: 3,
+            mediaRegistered: 0,
+            mediaFromSidecar: 0,
+            mediaMissing: 0,
+            mediaQuarantined: 0,
+            mediaDanglingRefs: 2,
+            mediaAmbiguous: 0,
+          })
+          expect(report.issues.filter((i) => i.includes('悬空'))).toHaveLength(2)
+        } finally {
+          await cleanupSiteRows(db, reg.siteId)
+        }
+      } finally {
+        await rm(work, { recursive: true, force: true })
+      }
+    })
+  })
+})
diff --git a/tests/unit/magic-bytes.test.ts b/tests/unit/magic-bytes.test.ts
new file mode 100644
index 0000000..71f2710
--- /dev/null
+++ b/tests/unit/magic-bytes.test.ts
@@ -0,0 +1,66 @@
+// magic-bytes 单测（Plan 3 Task 4 Step 1）：六类签名各正例 + 伪签名 null 例 +
+// SVG 大小写不敏感 + 窗口边界（1KB 外不检出）+ 二进制签名优先于文本特征。
+// 期望值全部手写字面量（与实现零共享——对拍而非引用）。
+import { describe, expect, it } from 'vitest'
+import { detectMediaType } from '../../apps/api/src/modules/import/magic-bytes.js'
+
+const PNG_SIG = Buffer.from([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a])
+
+describe('detectMediaType 手写签名表', () => {
+  it('六类二进制签名正例：PNG/JPEG/GIF87a/GIF89a/WEBP/PDF', () => {
+    expect(detectMediaType(Buffer.concat([PNG_SIG, Buffer.from('rest-of-png')]))).toStrictEqual({
+      type: 'image/png',
+      svg: false,
+    })
+    expect(detectMediaType(Buffer.from([0xff, 0xd8, 0xff, 0xe0, 0x00, 0x10]))).toStrictEqual({
+      type: 'image/jpeg',
+      svg: false,
+    })
+    expect(detectMediaType(Buffer.from('GIF87a-more-gif'))).toStrictEqual({ type: 'image/gif', svg: false })
+    expect(detectMediaType(Buffer.from('GIF89a-more-gif'))).toStrictEqual({ type: 'image/gif', svg: false })
+    expect(
+      detectMediaType(Buffer.concat([Buffer.from('RIFF'), Buffer.from([0x24, 0x00, 0x00, 0x00]), Buffer.from('WEBPVP8 ')])),
+    ).toStrictEqual({ type: 'image/webp', svg: false })
+    expect(detectMediaType(Buffer.from('%PDF-1.7\n%fake-payload'))).toStrictEqual({
+      type: 'application/pdf',
+      svg: false,
+    })
+  })
+
+  it('SVG 文本特征：大小写不敏感（<svg / <SVG / <Svg，含 xml 声明前导），svg=true', () => {
+    expect(detectMediaType(Buffer.from('<svg xmlns="http://www.w3.org/2000/svg"/>'))).toStrictEqual({
+      type: 'image/svg+xml',
+      svg: true,
+    })
+    expect(detectMediaType(Buffer.from('<?xml version="1.0"?><SVG onload="x()">'))).toStrictEqual({
+      type: 'image/svg+xml',
+      svg: true,
+    })
+    expect(detectMediaType(Buffer.from('<!-- c --><Svg width="1">'))).toStrictEqual({
+      type: 'image/svg+xml',
+      svg: true,
+    })
+  })
+
+  it('伪签名 / 纯文本 / 空缓冲 → null（未知签名不猜格式）', () => {
+    expect(detectMediaType(Buffer.from('hello plain text note'))).toBe(null)
+    expect(detectMediaType(Buffer.from([0x00, 0x01, 0x02, 0x03]))).toBe(null)
+    expect(detectMediaType(Buffer.alloc(0))).toBe(null)
+  })
+
+  it('RIFF 非 WEBP（WAV）不误报 webp → null', () => {
+    expect(
+      detectMediaType(Buffer.concat([Buffer.from('RIFF'), Buffer.from([0x24, 0x00, 0x00, 0x00]), Buffer.from('WAVEfmt ')])),
+    ).toBe(null)
+  })
+
+  it('二进制签名优先：PNG 头 + 尾部 <svg 片段仍判 PNG（文本窗口只对非二进制生效）', () => {
+    const sneaky = Buffer.concat([PNG_SIG, Buffer.from('<svg onload="alert(1)"> trailing')])
+    expect(detectMediaType(sneaky)).toStrictEqual({ type: 'image/png', svg: false })
+  })
+
+  it('SVG 窗口 1KB 钉死：<svg 出现在 1KB 之后不检出 → null', () => {
+    const beyond = Buffer.concat([Buffer.alloc(1200, 0x20), Buffer.from('<svg>')])
+    expect(detectMediaType(beyond)).toBe(null)
+  })
+})