exec.sh performance on Windows (Git Bash): where the time went #584

Field report, not a bug list. One small Stage 5 project (3 MDL scripts so far), Windows 11
Enterprise, Git Bash, mxcli v0.28.0, Mendix 11.12.6, toolkit c28fc0d. No project content quoted.

Headline: a script whose mxcli check takes 1.3 s and whose write takes 1-2 s costs 64-227 s
per exec.sh run. The model work is under 5% of that. Most of the rest is process-spawn overhead
in the bash helpers, multiplied by a slow spawn on this machine, plus two cold mxbuild runs.

Measurements

The ledger is .claude/loop/metrics-ledger.tsv (9 runs). Per-step times come from BUILD-LOG.

Machine baseline: a trivial process spawn (/usr/bin/true) costs ~94 ms. mxcli --version
costs ~420 ms. Defender real-time protection is on. Network, proxy and tunnel were all ruled out:
local disk, no VPN or tunnel active, and no network calls on the exec.sh path. Corporate Windows
laptops with endpoint protection commonly look like this, so it is probably typical for toolkit
users and not an outlier.

Findings, ranked by time saved

1. mxtk_ensure_mxbuild costs 7.2 s per run. It resolves the model version and the mxbuild
   path through many subprocesses, and repeats the work on every exec. Setting MXBUILD_PATH
   (in .claude/settings.local.json env) cut it to 1.6 s. The remainder is
   mxtk_model_version, still called for the mismatch warning.
   Suggest: memoize the model version and the mxbuild path per exec (or cache them keyed on
   the .mpr mtime), and have preflight recommend MXBUILD_PATH on Windows.
2. Cold mxbuild costs ~150 s, once per session, in baseline or gate. It is unavoidable, but
   it is invisible: the run looks hung.
   Suggest: print "first mxbuild of this session, expect ~2-3 min", or warm mxbuild in the
   background at session start.
3. Uncommitted-model refusal: 4 of 9 runs were blocked (8-19 s each, plus the human round
   trip). The second-exec idempotence check requires a commit in between, and agents keep running
   into this.
   Suggest: make the refusal message say exactly which commit to make, or allow the immediate
   idempotence re-run against the snapshot exec.sh just took.
4. Helper-script spawn count. guards (9 s), snapshot (5 s), score (5 s) and
   script-owners.sh (2.8 s) are almost all spawn overhead at ~94 ms each.
   Suggest: collapse the grep | sed | awk chains into single awk/bash builtins on the hot
   path; add MXTK_PROFILE=1 to print the spawn count and time per step so users can see this
   themselves.
5. The lint gate exits 2 on every run with no baseline (ERROR-NO-BASELINE, 0 errors). This
   costs no time, but it reads like a failure and sent investigation the wrong way.
   Suggest: state "no baseline yet: run bin/lint-gate.sh --update-baseline" in the exec.sh
   summary.
6. Preflight could detect slow spawn. Time 20 spawns. If the average is over ~30 ms, warn
   and point to the two real fixes: Defender exclusions for the project dir, Git's usr/bin and
   the JDK, or running inside WSL2 / the devcontainer.

Smaller friction hit in the same session

• Playwright is not installed by the toolkit, so the journey runner fails on first use.
• The journey runner guesses the app port and admin password wrongly. It should read them from
  the run config.
• check-design-reaches-app.sh has an argument-handling bug.
• The dotted row-ID regex rejects valid IDs.