[witbrock]

Merge decision: ready — unread navigation seeks directly to a bounded window instead of loading the entire exchange, preserving scoped access and history navigation.

User outcome

“Jump to first unread” no longer walks every historical page. In the matched 1,000-message browser fixture it falls from 2.08–2.62 seconds (19 requests, 1,000 rendered messages) to 66–77 ms (one request, 50 rendered messages). Native task: #V#task_agent_43536ba9d1dd211b359a0cf4709f8707.

Material changes

The existing exchange endpoint locates the earliest unread using the same actor/group/organisation/deletion scope, then returns a bounded chronological window with earlier/later cursors. The pane preserves pending position and retry, suppresses disconnected latest-page polling while reading an earlier window, and supports explicit return to Latest message. No schema/index migration or represented write.

Evidence

• 24 backend exchange/route tests pass, including authenticated actor routing, visibility, unread gaps, equal-time and both-direction pagination.
• 26 frontend exchange tests pass; changed-module static lint passes.
• Production frontend/CSS Playwright fixtures pass at 390/1280 px: unread navigation/read receipts and the matched latency comparison.
• Full scope, source-image hash and timings: docs/engineering/unread_message_latency_2026_09_17.md.

Ship boundary

Minimum ship criteria are removal of full-history traversal, correct chronological target and preserved bounded navigation/read state. Wrong actor/group visibility, skipped message ranges or broken retry would block shipment; tests found none.

Public database/network latency remains unmeasured: these browser timings use synthetic HTTP responses with 50 ms latency. There is no prepared live browser profile bound to this checkout. This limits the public under-one-second claim, not the demonstrated client traversal improvement. No deployment is requested or included. No secrets or runtime data committed.