[Performance] Low-end hardware performance contract #4

Goal

Treat shell, prompt, and editor latency as a compatibility constraint. The setup should remain quiet, direct, and usable on older, low-power, low-memory, or high-latency machines without trading away correctness, security, privacy, accessibility, or maintainability.

The reference machine is a ThinkPad X230 with a dual-core CPU, SATA SSD, and 2 GiB of available memory. Results from faster machines are useful observations, not substitutes for that baseline.

Principles

• Core shell and editor startup must not require a network connection.
• Startup paths must not perform automatic update checks or unbounded retries.
• Idle core shell/editor configuration must not poll or create background workers.
• X11 session services such as Picom, xss-lock, nm-applet, and dex are intentional profile behavior and are outside the zero-background-process goal.
• Performance changes require a profile or benchmark result; do not rewrite components from intuition alone.
• A numerical improvement never justifies weaker safety, rollback, provenance, or accessibility.

Measurement contract

• Use a clean temporary HOME and XDG tree.
• Separate runtime work from downloads and first-time provisioning.
• Report active optional host integrations when measuring configured Fish rather than treating them as isolated inputs.
• Measure Fish without config, configured interactive Fish, prompt rendering outside Git, prompt rendering in a small Git repository, and prompt rendering in a representative large repository.
• Measure warm Neovim startup only from an already-provisioned, lock-matching data tree and disable every provisioner in benchmark mode. Track first provisioning separately.
• Record commit, worktree state, hardware, operating system, runtime versions, load, warm-ups, and sample counts.
• Use at least 100 measured samples before reporting p99. Smaller runs must omit unsupported tail percentiles.
• Compare measurements only on similar hardware, power, thermal, and load conditions.

Provisional investigation thresholds

These values are hypotheses until the reference-machine baseline exists. They are not GitHub-hosted CI gates.

CI policy

GitHub-hosted runners validate behavior, portability, and bounded failure handling. Pull requests use offline Linux/macOS checks. Dependency installation stays under a 30-minute job ceiling, while the repository check step has a five-minute pull-request ceiling. Network provisioning runs only on main updates or manual dispatch, and the full Neovim cold graph runs only on Linux. Hosted-runner timing is too noisy for absolute startup latency gates; fixed-hardware measurements remain local and explicit.

Current workstation observation

Commit b6e23c7 on an Apple M4 running macOS 15.7.3, with five warm-ups and 100 measured samples:

Active integrations were eza, lolcat, and zoxide. This run is healthy and comfortably below the provisional prompt investigation threshold, but it is not the X230 acceptance baseline. Warm Neovim was intentionally skipped because a dedicated provisioned data tree was not supplied.

Progress

• Remove automatic update checks and shell-startup network work.
• Bound retries, locks, provisioning, rollback, and CI cancellation.
• Add scripts/benchmark with controlled Fish and prompt measurements.
• Consolidate prompt branch and dirty-state detection into one Git query.
• Record a current-workstation observation in Build a reliable personal dotfiles setup #5.
• Record the first baseline on the ThinkPad X230.
• Add a representative large-repository prompt scenario.
• Record warm Neovim measurements from an isolated provisioned data tree.
• Revisit the provisional thresholds using measured evidence.

This issue is intentionally incremental. Each performance change should be a small, independently verifiable improvement.