Issue 40278 Windows performance evidence
Base: 0dc32b5c8bb3c3408ef517d512bb17d9a96bedb0
Head: 9377e334187b9c9fa9143f1a0d89b090ea9d22a8

Environment: same leased Windows Server 2025 machine, 4 logical CPUs, about
16 GB RAM, Defender real-time protection and antivirus enabled. Production web
builds from separate installed checkouts; Node server from source. Chromium
headless, 1280x800, default light theme. No providers or replay turns needed.

Seed: local bare clone and local working clone with 33,228 tracked files.
Synthetic git identity, feature-0, one seed commit, origin set to the local bare
clone, fetched and pushed with upstream. Before each app revision/comparison,
feature-0 was reset to the seed commit and its bare remote branch force-reset.
Each side used its own fresh T3 home with seeding disabled and project add.
The bare remote was reused; earlier synthetic first-publish refs remained.

Driver: driver.cjs is the exact driver run on both revisions. It derives from
the issue's gitact2.cjs, adds video and a timing overlay, retains action results,
and checks clean status, upstream HEAD, and actual bare remote HEAD after every
run. Requests use the real paired page's app WebSocket, git.runStackedAction.
Timings start when the action Request is sent and end on its Exit, matching
the issue. They exclude preparation, checks and pauses. The driver does not
physically click the UI button. This is the same measurement path as triage.
Fresh single-use pairing token per scenario invocation. Six actions per scenario.

Order on each side: commit_push, push with upstream, first publish. Three edited
files and an explicit message for commit_push; one precommitted README edit for
push; a new branch on every first-publish run. Four-second pause before each
request and six-second pause after it, as in the issue's driver.

Two complete comparisons, 72 actions total. The first comparison improved all
medians, but push and commit_push improvements did not exceed full run ranges.
A second complete comparison was declared before it ran. Its three median
improvements exceed both sides' individual max-minus-min ranges. All samples
from both comparisons are retained; no run was discarded or retried.

Each initial/ and confirmation/ directory has six raw driver result JSON files,
two trace summaries, two action-span NDJSON files, and calculated measurement
JSON. Trace summaries group actions in execution order: indices 0-5 commit_push,
6-11 push, 12-17 publish. Full status counts are descendant statusDetails spans;
git counts are descendant runGitCommand spans. Timing-only NDJSON retains the
original action span IDs, parents, names, start timestamps and durations. Paths,
attributes, tokens and unrelated log messages are intentionally omitted.
extract.cjs produces those timing-only action traces; analyze.cjs reproduces the
counts; report.py produces the measurement images. reproduce.ps1 documents the
same seed and measurement sequence with caller-supplied paths.

Every action succeeded, left a clean worktree, and had local HEAD == upstream ==
actual bare remote HEAD. Returned commit/push statuses, upstream setup flags,
and Create change request CTA matched across revisions in every scenario.
Recordings retain the full six-run sequence; only the environment-name header
is covered with 'Windows test'. No speed changes or run edits were applied.

Scope limits: no skipped-up-to-date, alternate push remote, gone upstream,
default-branch/View PR toast, or post-push status-failure reproduction. Head
still exceeded the issue's absolute latency targets and some push Git-count
targets on this lease, despite the demonstrated relative improvement and 3-to-1
full-status reduction. This evidence is not a claim that all targets were met.
