diff --git a/NucleicRemote/NucleicRemote/Views/SessionDetailView.swift b/NucleicRemote/NucleicRemote/Views/SessionDetailView.swift
index e5f3b16..82482a8 100644
--- a/NucleicRemote/NucleicRemote/Views/SessionDetailView.swift
+++ b/NucleicRemote/NucleicRemote/Views/SessionDetailView.swift
@@ -494,37 +494,15 @@ struct TranscriptList: View {
     /// settling — a `.task(id:)` debounces it to detect when the opening layout has come to rest.
     @State private var transcriptContentHeight: CGFloat = 0
 
-    /// Memoized transcript projection. `TranscriptProjection.build` folds the entire raw event
-    /// stream into rows on every read, but `body` re-evaluates far more often than `events`
-    /// actually changes: while the chat is opening, `onScrollGeometryChange` bumps
-    /// `transcriptContentHeight` on every settling-layout frame, and each bump re-runs `body` —
-    /// so the untouched event stream was re-folded dozens of times during the exact window the
-    /// tab bar is sliding away, starving that animation of the main thread and making the
-    /// load-in jitter. Reuse the last fold whenever the stream is unchanged. A class held in
-    /// `@State` so reading/updating it from `body` doesn't itself invalidate the view (SwiftUI
-    /// stores the reference and never diffs its interior).
-    private final class ProjectionCache {
-        /// The event stream is append-only with strictly increasing `seq`, so `(count, lastSeq)`
-        /// uniquely pins it; `firstSeq` guards the open reset (stream cleared to `[]`, then
-        /// refilled). The two display toggles are inputs to the fold, so they key it too.
-        private struct Key: Equatable {
-            var count: Int; var firstSeq: UInt64; var lastSeq: UInt64
-            var showRaw: Bool; var showLockEvents: Bool
-        }
-        private var key: Key?
-        private var value: [TranscriptItem] = []
-
-        func items(for events: [AgentEvent], showRaw: Bool, showLockEvents: Bool) -> [TranscriptItem] {
-            let k = Key(count: events.count, firstSeq: events.first?.seq ?? 0,
-                        lastSeq: events.last?.seq ?? 0, showRaw: showRaw, showLockEvents: showLockEvents)
-            if key == k { return value }
-            let built = TranscriptProjection.build(events, showRaw: showRaw, showLockEvents: showLockEvents)
-            key = k
-            value = built
-            return built
-        }
-    }
-    @State private var projectionCache = ProjectionCache()
+    /// Incremental transcript projection (docs/TRANSCRIPT_INCREMENTAL_PROJECTION.md). Two layers:
+    /// a read memo that collapses the redundant `body` re-evaluations (settling layout bumps
+    /// `transcriptContentHeight` every frame while opening, and each bump re-runs `body` against
+    /// an unchanged stream), and — when the stream *has* grown — a stable-prefix fold that seals
+    /// everything before the live turn once and re-folds only the tail, so a streaming delta
+    /// costs O(live-tail) instead of re-folding all N events (which made long sessions cost
+    /// O(N²) over their lifetime). A class held in `@State` so reading/updating it from `body`
+    /// doesn't itself invalidate the view (SwiftUI stores the reference, never diffs interior).
+    @State private var projectionCache = IncrementalTranscriptProjection()
 
     /// The in-flight background Markdown pre-warm (below), cancelled when a newer one supersedes
     /// it or the transcript goes away — so a long warm can't outlive the view or stack up behind