perf(test): remove redundant serial PTY idle waits - #1276

Summary

Optimize the existing single-runner, serial PTY suite. The sharding experiment has been removed completely: no workflow changes, matrix, extra runners, or concurrent PTYs remain in this PR. Other jobs, release branches/tags, and release workflows are untouched.

• Replace redundant shell-completion idle waits with a unique fixture acknowledgement emitted only after each captured-arguments file closes. Send input/Tab/Enter in Readline order and retain every Bash/Zsh quoting, Unicode, option-boundary, and loading-mode assertion.
• Deliver intermediate pointer movements without 500ms waits when they cause no repaint. Still observe real hover controls, selection, and saved-card state.
• Replace composer/paging sleeps with code, keyboard-readiness, and full-footer predicates.
• Carry the PTY hunks from coordinator commit 0d16237f forward: chore(release): prepare v0.25.0 #1274 merged without that commit. These cover CSS palette menu closure, CJK reflow, wheel/header layout, exact-line jitter, feedback readiness, and isolated default working directories. The Windows metadata fix is outside this PR's scope.
• Empty maintenance Changeset and profiling/maintenance notes in test/pty/README.md.

Controlled real CI measurements

Both runs are on this pushed branch, the same 62064bcf source base plus the same coordinator stabilization fixes. No retries, no difference in job definition or test inventory:

• Before: 43053286536, b720f828, single-runner control before optimization.
• After: 43054778196, 1575d0c1, three-file optimization plus measurement docs.

Job durations are completed_at - started_at from gh api repos/modem-dev/hunk/actions/runs/<id>/jobs, including setup and cleanup.

Observed PTY reduction: 72s / 16.3%. PR workflow wall time: 7m52s → 6m29s. PTY remains the critical path, now close to Windows; no claim that the untouched jobs improved.

From individual test-duration sums in the real PTY job logs:

About 34 seconds disappeared directly from the modified files. Other file durations also varied, so the entire observed 72-second job improvement is not promised on every runner. Both jobs executed all 291 tests across 41 files: 287 pass, 4 identical existing environment skips, 1,214 assertions. No functional assertions or integration tests were deleted. The prior sharding-only helper/tests were removed with that rejected experiment, not from the original suite.

Confirmation on the final documentation head 43a59868: 43056271912, all PR CI workflow jobs green, no retries. The single PTY job took 6m20s (14.2% below the 7m23s control). Untouched jobs: Linux 3m19s, Windows 5m33s, compiled headless Windows/macOS 46s/23s, changeset 9s, validation gate 2s, change detection 5s. The observed serial PTY range across the two after runs is 6m11s–6m20s, not the superseded matrix timings.

Validation and rejected alternatives

• bun run test:integration before pushing: 290 pass, 1 unchanged platform skip, 1,252 assertions, 339.57s locally, versus a 391.19s baseline on the refreshed source base (13.2% less local time). CI lacks local zsh/Neovim, explaining its different pass/assertion counts.
• Four local completion tests: 50.90s → 3.94s, same 76 assertions.
• Threaded-note test: ~7.9s → 4.08s, same 15 assertions.
• bun run typecheck, targeted oxlint, formatting, and git diff --check: pass. Controlled before/after PR CI workflows both passed, including Windows and Linux validation.
• A precompiled executable did not improve the sampled notes/navigation/file-view tests (64.38s versus ~62.23s from source); no build/artifact dependency added.
• A broad event-only screen-wait rewrite exposed a multi-frame resize/reveal race and was discarded. General screen-settling behavior is unchanged.
• Keep the deliberate 62-second suspension check, fixed wheel/drag cadence, and negative/stability checks where timing is part of the tested behavior. Builds on perf(test): reduce fixture process overhead #1092 without repeating its fixture-process refactor.

Explicit performance-policy change and risks

The coordinator's dialog fix replaces the noisy cold-minus-warm differential assertion with independent absolute cold/warm gates. Cold limits remain 150ms idle / 200ms first-frame, and warm reopen has an explicit 150ms limit. A very fast warm reopen no longer invalidates an otherwise in-budget cold open. No timeout is blindly increased; this is the existing coordinator fix, called out rather than hidden as a speed optimization.

Acknowledgements belong to the shell fixture and are emitted after closing the captured file; they do not substitute dispatch ACKs for Hunk screen assertions. Faster pointer delivery relies on ordered SGR events, followed by destination-state checks. Shared-runner timing varies; the PR reports both targeted-file gains and full-job observations, not a guaranteed speedup. All original platform gates and coverage remain.

This PR description was generated by Pi using GPT-6.1-sol