Batch AI events during conversation restoration.

[vorporeal]

Description

Viewing a large cloud agent conversation could beachball the UI for ~30s. When a viewer joins a shared session, the event loop replays the entire event backlog on the main thread, and every replayed message delta emits a BlocklistAIHistoryModel event that fans out to ~22 subscriber types via flush_effects. Profiling showed ~92% of the freeze was this per-delta subscriber fanout. This got much more noticeable recently because opening a running cloud conversation now attaches to the live session (#11097) and orchestrated sessions are shareable (#11465), so the replay path runs far more often.

This PR coalesces history-model events during viewer catch-up so subscribers see one fanout for the whole backlog instead of one per replayed delta:

• warpui_core: adds ModelContext::capture_emitted_events_for_model, which runs a closure and extracts events queued for a target model instead of leaving them for delivery.
• BlocklistAIController: while a coalescing window is active (event-backlog catch-up or an existing-conversation replay), shared-session response handling captures history events into a buffer with last-write-wins dedupe for high-volume events (UpdatedStreamingExchange per exchange, todo list, conversation status/metadata/usage, active conversation). All other events are preserved in order. The buffer is flushed as one batch when the window ends.
• Viewer EventLoop: begins coalescing when joining with a backlog, tracks AgentConversationReplayStarted/Ended brackets, and flushes at catch-up completion.

Deferring is safe because events are already delivered asynchronously via flush_effects, payloads carry IDs that handlers use to re-read current model state, and the deduped variants' snapshot fields keep the latest payload.

The behavior is gated behind FeatureFlag::CoalesceSharedSessionCatchUpEvents (dogfood-only for now) so we can validate internally before rolling it out.

Plan: https://staging.warp.dev/drive/notebook/k3sdmZpC3ZnRo0SJz6iLAT
Conversation: https://staging.warp.dev/conversation/f0493bea-7990-4f61-9e46-bb9cb3f559ce

Testing

• I have manually tested my changes locally with ./script/run
• Manually opened a large running cloud agent conversation as a viewer: no extended beachball during catch-up, transcript renders correctly once caught up, and live streaming updates continue normally afterwards.
• cargo check, ./script/format, and cargo clippy --workspace --all-targets --all-features --tests -- -D warnings pass.
• cargo nextest run -p warp -E 'test(event_loop) or test(shared_session)': 175 tests pass.

Agent Mode

• Warp Agent Mode - This PR was created via Warp's AI Agent Mode

Co-Authored-By: Oz oz-agent@warp.dev

[vorporeal]

• Batch AI events during conversation restoration. #12597 👈 (View in Graphite)
• master

This stack of pull requests is managed by Graphite. Learn more about stacking.