[nvdorman]

Summary

• stop re-downloading and re-hydrating the entire transcript on every idle chat attachment poll
• coalesce text/reasoning deltas in the browser patch queue and group patches per message
• coalesce replay backlog server-side and expose absolute cursors so reconnects do not replay or duplicate old deltas
• add frontend regression tests to make check

Root cause

The backend remained healthy while the browser renderer stalled. Idle /chat/attach polls returned done, and each response fetched and rebuilt the full transcript every 1.5 seconds. Reconnects also always used cursor=0, replaying up to 4,000 token-sized events. Long sessions therefore repeatedly discarded memoized message objects and performed thousands of growing-string operations on the main thread.

Verification

• make check
• CHROME=/opt/brave-bin/brave make smoke (15 routes clean)
• isolated runtime against a snapshot of a 131-message / ~205 KiB session: title rendered, 399 DOM nodes, 0 exceptions, transcript detail requests stopped after initial load + sidebar + one race-closing refresh
• regression: 4,001 retained live-run events replay in 2 frames with correct absolute cursors
• frontend regression: 4,000 adjacent deltas coalesce to 1 queue item