Never double-import a file into the engine: detail Transcribe bails when the pump already owns the file (in-flight check was absent there — mid-upload the mapping sidecar doesn't exist yet, so the upload path created a second server recording; happened live with Gambino Tahoe 11-50). Pump's transcribeFile reuses an existing mapping via /reprocess instead of uploading fresh (a pipeline killed mid-run no longer re-imports as a new recording on next pump). Purged the 11:08 orphan recording from the engine DB.

app/src/main/kotlin/com/shonar/desktop/DesktopState.kt

@@ -673,6 +673,19 @@ class DesktopState(private val appDir: File = defaultAppDir()) {
             return false
         }
         return try {
+            // If an earlier pass ever uploaded this file, reuse that
+            // recording and re-run the pipeline in place. Uploading again
+            // would create a duplicate server recording (this is exactly
+            // how a file whose pipeline died mid-run got re-imported as a
+            // second row on the next pump).
+            val existing = loadMapping(f)?.recordingId
+            val refKey: String
+            if (existing != null) {
+                refKey = existing
+                runCatching {
+                    provider.reprocess(existing, "transcribe")
+                } // 409 "already running" is fine: we adopt it by polling.
+            } else {
                 val draft = RecordingDraft(
                     id = UUID.randomUUID().toString(),
                     title = f.nameWithoutExtension,
@@ -685,12 +698,14 @@ class DesktopState(private val appDir: File = defaultAppDir()) {
                 )
                 val ref = provider.upload(draft) {}
                 saveMapping(f, RemoteMapping(ref.key))
+                refKey = ref.key
+            }
             var jobs: List<JobInfo> = emptyList()
             var polls = 0
             while (polls < 600) { // up to ~30 min per file at 2s polls
                 polls++
                 delay(2000)
-                jobs = runCatching { parseJobs(provider.fetchJobs(ref.key)) }.getOrNull().orEmpty()
+                jobs = runCatching { parseJobs(provider.fetchJobs(refKey)) }.getOrNull().orEmpty()
                 setLive(f.name, jobProgress(jobs))
                 rescanStatuses()
                 val t = jobs.firstOrNull { it.jobType == "transcribe" }?.status
@@ -702,9 +717,9 @@ class DesktopState(private val appDir: File = defaultAppDir()) {
                 it.jobType == "transcribe" && it.status == "failed"
             }
             if (transcribeFailed) return false
-            val transcript = runCatching { parseTranscript(provider.fetchTranscript(ref.key)) }
+            val transcript = runCatching { parseTranscript(provider.fetchTranscript(refKey)) }
                 .getOrNull() ?: return false
-            val summary = runCatching { parseSummary(provider.fetchSummary(ref.key)) }.getOrNull()
+            val summary = runCatching { parseSummary(provider.fetchSummary(refKey)) }.getOrNull()
             saveReport(DetailUi(file = f, transcript = transcript, summary = summary))
             true
         } catch (e: Exception) {
@@ -1377,6 +1392,17 @@ class DesktopState(private val appDir: File = defaultAppDir()) {
     fun transcribe() {
         val d = _detail.value ?: return
         pollJob?.cancel()
+        // The pump (or a detail-run from elsewhere) already owns this
+        // file: never start a second pipeline for it. The upload path
+        // below creates a NEW server recording, and the mapping sidecar
+        // only exists once an upload completes — so mid-flight this is
+        // the only reliable dedup. Show the live row state instead.
+        if (d.file.name in inFlight || isRunning(d.file.name)) {
+            d.busy = _liveProgress.value[d.file.name]?.label ?: "Already queued…"
+            d.error = null
+            _detail.value = d.copy()
+            return
+        }
         // A file with a known server recording re-runs the pipeline IN
         // PLACE (the endpoint takes the model override too) — re-uploading
         // would duplicate the recording server-side. Only never-uploaded