perf(terminal): silence replies provoked by replay and coalesce keystrokes

[Juliusolsson05]

Problem

Attaching an inline terminal writes up to 512 KiB of replay into xterm in one write. xterm answers every terminal query embedded in that stream — cursor-position reports (ESC[6n), device attributes, OSC colour queries — through term.onData, and the three xterm hosts forwarded each chunk to the provider with its own ipcRenderer.invoke. Today's performance journal: one attach at 20:18:57 → 21,795 session:input calls in 15 s and a 7,448 ms renderer long task; other attaches → 2,071 and 2,236 calls; a steady ~5 Hz trickle while any pane is attached. Main then wrote thousands of stale ESC[row;colR replies into the agent's stdin — answers to questions the provider asked minutes ago about a screen it has since repainted. Details in #745.

What changed

• New src/renderer/src/workspace/tile-tree/terminalInputForwarder.ts: replay(term, chunks) writes the replay and holds a latch until xterm's write callback confirms each chunk is parsed — everything onData emits meanwhile is dropped; onData(data) outside a replay coalesces same-tick chunks into one sendInput (one microtask of latency).
• AgentTerminalLeaf, TerminalLeaf and AgentInlineTerminal use it for both the buffer replay and the pre-attach backlog, and acknowledge user activity only for data that was actually forwarded (a replay reply is not interaction).

Keystrokes typed during the ~100 ms replay window are dropped with the replies — the pane has not painted yet, and a stale reply cannot be told from a real key (both are escape sequences).

Linked issues

Fixes #745. Refs #103, #390.

Verification

• terminalInputForwarder.test.ts (unit): replies emitted while a replay is parsed are dropped and nothing is sent; live data after the replay is forwarded, coalesced per tick in order; an empty replay resolves without holding the latch; multi-chunk replay keeps order and releases only after the last callback. 4 tests passed.
• --project renderer src/renderer/src/workspace/tile-tree/ — passing; npx tsc -b — clean.
• After: session:input bursts on session:agent-pty-attach should disappear from the journal and the renderer long task with them.

Not in scope

Answering DSR/DA in the headless terminal so responses are consistent whether or not a pane is attached, and switching keystroke transport to a fire-and-forget send (both noted in #745).

Plan: docs/superpowers/plans/2026-09-03-terminal-replay-input-storm.md.

Known limitations

• Replies to the TUI's LIVE queries are still forwarded one session:input per repaint burst (~5 Hz per attached pane) and still call acknowledgeSession, which clears the unread/ACTION badge without user interaction. Tracked in perf(terminal): live TUI query replies still cost one session:input each and clear the unread badge #751; the structural fix is answering DSR/DA in the headless terminal so the renderer never has to.
• Keystrokes typed during the replay parse window are dropped with the replies (the pane has not painted yet). The window is one task hop plus the synchronous parse of up to 512 KiB, unmeasured; worth a journal number post-merge.
• A TUI whose startup probes land in the pre-attach buffer (global terminal view mounts the pane at spawn) no longer gets them answered on replay and waits out its probe timeout — the same behaviour feed view and background sessions already had, now consistent.
• Two consumers on one session (tile leaf plus the debug inline terminal) still answer each live query twice (pre-existing).

🤖 Generated with Claude Code

https://claude.ai/code/session_01Kk16MNVJtWAnCeGxGRuHqa

[Juliusolsson05]

Review applied:

• Host-level guard: the leaf test's xterm mock now invokes the write callback, and a new case types a key after attach and asserts it reaches sendInput (and acknowledges) — the deaf-pane failure mode is now covered through the real wiring.
• Synchronous callback path has its own forwarder test.
• Throwing term.write releases the unqueued chunks' share of the latch and rejects instead of leaving the pane deaf.
• Live-reply trickle + acknowledge clearing the ACTION badge split into perf(terminal): live TUI query replies still cost one session:input each and clear the unread badge #751 (this PR keeps Fixes #745 for the replay storm it measured); PR body gained a Known limitations section incl. the startup-probe consistency note; paste wording corrected; inline terminal joins its backlog like the leaves.