perf(screen): session:screen frames are 50% duplicate strings and 99% spinner-only diffs #746

Symptom

session:screen is the highest-volume, highest-byte IPC channel and almost all of it is redundant.

Evidence (session recording 794dafdd, 13.2 min, one Claude session): 4,728 frames = 6 Hz, avg 8.3 KB/frame (p99 13.4 KB). Per frame: plain 2,109 B, recent 2,109 B, markdown 1,853 B, recentMarkdown 1,853 B — recent === plain in 4,728/4,728 frames and recentMarkdown === markdown in 4,728/4,728 (an alt-screen TUI has no scrollback). Consecutive frames: 78% differ in exactly one line and 21% in two — the spinner line (✻ Beboppin'… → ✽ Ionizing… → ✶ … (5s · thinking…)). 0% byte-identical, so the renderer's no-op bail never fires. This run: 21,700 frames, 285 MB recorded; one session 8,913 frames = 97 MB.

Mechanism

packages/claude-code-headless/src/terminal/HeadlessTerminal.ts:600-614 gates on exact equality of plain/recent, then computes both markdown walks and emits four strings every 100 ms; ClaudeCodeHeadless.ts:614 forwards every snapshot; src/providers/claude/runtime/claudeSession.ts:490-497 and src/main/sessions/forwarder.ts:57 send them through the 100 ms latest-wins coalescer; useIpcSubscriptions.ts:739-840 commits a new runtime object per frame (chrome-tick frames skip the feed-debug append but still re-render the workspace at 6 Hz × active sessions); SessionRecorderManager JSON-stringifies each frame.

Expected

• Where recent === plain (and recentMarkdown === markdown), send the field once and let the renderer alias it — halves bytes and removes one markdown walk per frame. Can be done at the src forward point without touching the package.
• Spinner-aware gate at the forward point: normalise the spinner glyph and the (Ns · …) timer before comparing to the last emitted frame; condition parsers keep running on the raw frame. Renderer: return prev for spinner-only diffs.
• markdown is debug-only per the comment in useIpcSubscriptions.ts:786-800; consider not sending it unless a debug consumer is mounted.

Impact

~10× fewer screen IPC messages while an agent is thinking, −50% bytes on the rest, half the markdown conversion CPU in main, ~−250 MB per hour of session-recording IO, far fewer workspace re-renders.

Refs #103, #390.