Publish the resource benchmark harness

[documentnode-dev]

TermTree's memory, CPU, and cold-start claims rest on a measurement harness. This publishes it, so anyone can run it against their own installed applications rather than take the numbers on trust. A benchmark nobody can run is an assertion, not evidence.

benchmark/ is the one deliberate exception to this repository's source-free posture, licensed Apache-2.0 under benchmark/LICENSE and scoped to that directory alone. LICENSE.md and README.md are updated to say so — the previous wording asserted the repository contained no source at all.

What it measures

On macOS, every off-the-shelf attribution method is biased, and in a direction that depends on the runtime under test. WebKit helper processes are launchd-parented, so a process-tree walk misses them entirely; LaunchServices, conversely, does not enumerate an Electron app's helper processes. Measuring either runtime with one mechanism alone undercounts it severalfold. The harness takes the union of both, and reports memory two ways rather than picking the flattering one.

Safety and portability

The harness seeds application state and launches real applications, so being safe for someone other than its authors to run is a precondition, not a nicety.

• Every subject is seeded, launched, and measured against a disposable per-run home directory under the OS temp dir. The runner's real application profiles are never read or written, and an explicit --home pointing at the real home directory, or an ancestor of it, is refused rather than obeyed. A pristine profile is also what makes independent runs comparable at all.
• Subjects launch via open -n -F --env HOME=<scratch>, which keeps a real LaunchServices registration so helper attribution still resolves.
• A run refuses to start when a subject is already running, keyed on bundle identifier rather than app name: two differently named bundles can declare the same identifier, and the second launch is then handed off to the running instance and exits having measured nothing.
• Stale assumptions fail loudly instead of degrading silently — unrecognised cold-start log marks, a missing app data directory, and an unconsumed seed each invalidate the sample by name.
• The Collaborator, CodeNomad, and diri seeders have never been checked against a real install. Their heavier tiers report invalidReason: "seed-format-unverified" rather than being counted as valid.
• Bundle paths, the seeded repo, and the agent CLI are all overridable, so installs outside /Applications work.
• doctor names every unmet prerequisite and how to clear it, and changes nothing.

Scope and honesty

macOS only, and structurally so: lsappinfo, footprint, vm_stat, notifyutil, and pmset have no Linux or Windows equivalents.

The README is explicit that the method and within-run ratios travel between machines while absolute figures do not — a reader on different hardware will not reproduce a given number and should not expect to.

The CodeNomad Electron-vs-Tauri pair is a supported standalone invocation: one MIT codebase, two runtimes, TermTree not a participant. Someone can run that alone to check the attribution correction behaves as claimed before believing any TermTree figure.

Verification

CI (new to this repository) runs the format check, clippy, and tests on macOS. It deliberately never runs a sweep — that needs an exclusive, quiesced machine for hours, and numbers from a shared CI runner would be misleading.

175 unit tests pass. benchmark/doc/resource-benchmark-harness-testing.md documents the manual pass, and records what is deliberately not verifiable without a quiesced machine — including launch isolation against a real application, which needs the subject quit first.

No measurement results are published here yet.