[simota]

A static performance scan found four places where cost grew without a bound on hot paths. This fixes them.

• Blank-line floods bypassed scrollback-limit. Rows that trim to zero cells were charged 0 bytes and never filled a page, so yes '' grew history and memory forever. Each row is now charged its RowMeta size, and pages also seal on a row-count cap.
• Busy background tabs cost O(K²) main-thread Terminal locks. Each Redraw from an occluded pane re-titled every dirty occluded window. Titles are now re-asserted only for windows leaving occlusion; still-occluded windows already refresh their own title on their own Redraw.
• Remote panes posted one main-thread Redraw per WebSocket frame. Each Redraw locks the terminal and rebuilds the session card. Remote frames now go through the local path's redraw floor (decide_redraw_floor) with a trailing redraw.
• Overtaken remote scrollback snapshots were refetched in a tight loop. Each refetch cost a connect, a server-side tail walk, and two full local scrollback builds. The first refetch stays immediate; later ones back off exponentially from 1s to 30s.

Reviewer notes:

• feed_remote_bytes no longer requests a redraw; run_connected decides it via remote_redraw_decision. Three existing tests asserted one redraw per fed batch; those assertions were removed because that is the behavior being changed.
• Under synchronized output, the trailing redraw is anchored to the latest batch (now + SYNCHRONIZED_OUTPUT_MAX_SUPPRESSION), as before this branch. Anchoring it to the last redraw request lets it land inside the render side's snapshot hold window, where it would be swallowed.
• Overtaken refetches are not capped. Raw output can overtake two fetches in a row and a later fetch still merges, so giving up would drop older history for the whole connection.
• Not addressed here: main-thread reflow of every pane on font-size, sidebar, or scale changes, and the O(rows × pages) PagedScrollback::row lookup. Both need a reflow-path redesign and are left for a follow-up.

Validation:

• cargo test --workspace --exclude noa-pty --exclude noa-ipc: 2,543 passed, 0 failed. noa-ipc tests need loopback bind and noa-pty tests need openpty; both were unavailable in the sandbox, and neither crate is changed.
• New regression tests: blank-row flood stays within the limit; remote redraws are floored during a 2,000-frame flood; the synchronized trailing redraw is anchored to the latest batch; overtaken refetch backs off instead of spinning; scrollback still merges after two overtaken fetches; backoff delay schedule.
• Each new regression test was confirmed to fail with its fix reverted and to pass with it applied.
• cargo clippy --workspace --all-targets shows no new warnings. rustfmt --check on changed files and git diff --check passed.

Limitations:

• Static analysis only; no profiling or GUI run. The size of the main-thread stalls was not measured.
• The occluded-tab title change has no App-level test harness and was verified by reading the code.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-14T01:41:00.898920Z	765bf21	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.