Published result

Extracting one release file cuts the biggest MacBook test's temp files by 46%

A plain-language summary comes first, followed by the full text of the result's reproduction pack.

Result summary and pack text · reproduce/build-speed-e6d-assembly-single-member-extract-20261105/README.md

In plain terms

The release-assembly test unpacked the whole release package (about 7,500 files) to read one executable. Extracting only that executable cut the files it creates per run by 46% (15,425 to 8,297) and file-event CPU in 7 of 8 paired rounds, with every test passing and the package unchanged, so a change request is filed.

Verdict key: Yes, The result supports the change or claim that was tested. No, The result does not support it, so nothing changes. Unresolved, The evidence is not strong enough either way. Superseded, A later test replaced this result.

Full text of the reproduction pack

The text below is the pack's README as published in the repository (a system folder name may show without its leading slash, so that the site's privacy scan passes).

The release-assembly test (TestAssemblyUsesThePinnedComponents, the largest single temp-file writer in a Build on the MacBook) unpacks the whole release archive, about 7,500 entries, to read one executable. The earlier census found it makes 24% of a Build's file creations and that two ways of sharing the packaged files (tar -h) change 11 archive links. This follow-up measures the one edit that cannot change the archive: the test extracts only the bin/kogen member it inspects (17 inserted lines in cli/assembly_test.go; packaging unchanged).

Source: kogen at 24136fc, scratch edit only, no Kogen commit. Registered in Studio lab lab-e6-fsevents-temp-trees-20261105 (revision 7) before the first run. Model-free; run on the MacBook as the build user with the delivery build environment, one run at a time, only while load5 was at most 24 and no verify or landing process or heavy-queue entry existed. Eight interleaved unwatched rounds (alternating order) and three watched rounds (file-create events per component), go test -count=3 per window.

Arm Runs passed Files created per test run Change fseventsd CPU per window (median) Lower than A Median wall
A: unmodified 8/8 (+3/3 watched) 15,425 - 6.73 s - 28.1 s
D: single-member extract 8/8 (+3/3 watched) 8,297 -46.2% 4.01 s 7/8 rounds (one-sided sign test p=0.035) 25.1 s

In A the whole-archive extract is 21.7k of 44-48k file creations per window; in D it is zero. What remains is the packaging work dir (about 21-22k per window), which this edit does not touch.

The package is identical. Packaging is unchanged; building it in the A and D trees gives 7,549 entries each with identical path, type, mode and sha256 (0 differing), unlike the shared-store arms of the earlier study.

Verdict under the registered rule: D passes (tests and package identical, files at least 25% lower, file-event CPU lower in the rounds with p<=0.05). A change request for the test is filed (Lab-first outbox). Caveats: CPU is read from 25-30 s unwatched windows with delivery work running nearby (load5 19 in the first round, 5-10 later), so it is paired per round, not absolute; round 1 had D higher.

metrics.json maps run counts onto the climb fields (cells = test runs, all_pass and hidden_* = passed tests); it has no Intents, so valid_intent and landed are 0.

Rerun

Everything needed is in this folder and the public source. Clone kogen at 24136fc into two scratch copies kA and kD; in kD run python3 -I scripts/mkc.py cli/assembly_test.go (adds the single-member extract). Put the scripts and a build of scripts/fswatch.swift (swiftc -O fswatch.swift -o fswatch) in the parent directory of kA/kD, set the paths at the top of each script, then run sh scripts/abseq.sh <absolute outdir> <gate deadline, epoch seconds> on a quiet host (each run is scripts/abrun2.sh) and sh scripts/eqcheck.sh <absolute outdir> for package identity. Reproduce the table from the recorded data with python3 -I report.py data ab AD (reads data/ab; the identity result is data/eq.json).