Drop stuck whole-call visual spans at processing time

Defense-in-depth + salvage for sessions captured before the adapter fix: drop any
vision-source span whose single unbroken duration covers ≥60% of the call. No one
speaks that long without a break, so it's a stuck/false active-speaker cue that
would dominate backend name attribution. Self (mic_vad) spans are never dropped.
Applied to both the live and re-process paths. Test added; 66 pass.

Ten31Transcripts/Session/TranscriptPipeline.swift

@@ -50,13 +50,20 @@ final class TranscriptPipeline {
         try? fm.createDirectory(at: chunksDir, withIntermediateDirectories: true)
         defer { try? fm.removeItem(at: chunksDir) }   // cleanup on success OR throw
 
+        // Defensive: drop any visual span covering most of the call in one unbroken
+        // segment — the signature of a stuck/false active-speaker cue (e.g. a solid
+        // camera-off tile read as "speaking" the whole call). Such a span would
+        // dominate the backend's name attribution and collapse every voice onto one
+        // name. Also salvages sessions captured before the adapter fix landed.
+        let vis = Self.dropStuckSpans(timeline, duration: duration)
+
         // Start from stored voiceprints; accumulate this call's prints across chunks
         // for within-call unification (the store only persists high-confidence ones).
         var known = voiceprints.knownVoiceprints()
         var results: [TranscriptAssembler.ChunkResult] = []
         // Mono fallback needs self folded into the timeline; dual sends it separately.
-        let monoTimeline = dual ? timeline
-            : timeline + Self.timeline(fromSelfSpans: selfSpans, selfName: selfName)
+        let monoTimeline = dual ? vis
+            : vis + Self.timeline(fromSelfSpans: selfSpans, selfName: selfName)
 
         for chunk in plan {
             try Task.checkCancellation()
@@ -70,7 +77,7 @@ final class TranscriptPipeline {
                 try SessionPackager.sliceAudio(from: micURL, startSec: chunk.start, endSec: chunk.end, to: micChunk)
                 try SessionPackager.sliceAudio(from: systemURL, startSec: chunk.start, endSec: chunk.end, to: sysChunk)
                 guard fm.fileExists(atPath: micChunk.path), fm.fileExists(atPath: sysChunk.path) else { continue }
-                let timelineData = try SessionPackager.rebasedTimelineData(timeline, start: chunk.start, end: chunk.end)
+                let timelineData = try SessionPackager.rebasedTimelineData(vis, start: chunk.start, end: chunk.end)
                 let selfVadData = try SessionPackager.rebasedSelfVadData(selfSpans, start: chunk.start, end: chunk.end)
                 response = try await client.labelMergeDual(
                     micURL: micChunk, systemURL: sysChunk, selfName: selfName, selfVad: selfVadData,
@@ -113,4 +120,14 @@ final class TranscriptPipeline {
     static func timeline(fromSelfSpans spans: [VADSpan], selfName: String) -> [VisualTimeline.Segment] {
         spans.map { .init(start: $0.start, end: $0.end, name: selfName, confidence: $0.confidence, source: "mic_vad") }
     }
+
+    /// Drop visual (vision-source) spans whose single unbroken duration covers at
+    /// least `maxFraction` of the whole call — no one legitimately speaks that long
+    /// without a break, so it's a stuck/false cue. Self spans (mic_vad) are kept.
+    static func dropStuckSpans(_ timeline: [VisualTimeline.Segment], duration: Double,
+                               maxFraction: Double = 0.6) -> [VisualTimeline.Segment] {
+        guard duration > 0 else { return timeline }
+        let limit = maxFraction * duration
+        return timeline.filter { $0.source != "vision" || ($0.end - $0.start) < limit }
+    }
 }

Ten31TranscriptsTests/Phase5Tests.swift

@@ -37,6 +37,18 @@ final class Phase5Tests: XCTestCase {
         XCTAssertEqual(asm.speakersFile.segments[0].start, 152, accuracy: 0.01)
     }
 
+    func testDropStuckSpansRemovesWholeCallCue() {
+        let segs = [
+            VisualTimeline.Segment(start: 0, end: 1900, name: "Grant Gilliam", confidence: 1, source: "vision"), // stuck whole-call tile
+            VisualTimeline.Segment(start: 100, end: 130, name: "Matt Odell", confidence: 0.9, source: "vision"),  // real
+            VisualTimeline.Segment(start: 0, end: 1900, name: "Grant", confidence: 1, source: "mic_vad"),         // self span: keep
+        ]
+        let out = TranscriptPipeline.dropStuckSpans(segs, duration: 1976)
+        XCTAssertFalse(out.contains { $0.name == "Grant Gilliam" })   // 96% of call in one span → dropped
+        XCTAssertTrue(out.contains { $0.name == "Matt Odell" })       // short real span kept
+        XCTAssertTrue(out.contains { $0.source == "mic_vad" })        // self never dropped
+    }
+
     func testRebaseClipsAndRebases() throws {
         let segs = [
             VisualTimeline.Segment(start: 140, end: 160, name: "A", confidence: 0.9, source: "vision"),