[MendixMau]

#583: Stage 3's gate-check plus closeout took about 12 min on a 1-module, 7-page slice. Profiling on Linux showed no slow step. The cost is about 2,000 small process starts per run; at about 100 ms each on Windows, that comes to minutes.

What changes

1. --closeout reuses the verdicts of the gate-check that just ran. The cache is keyed on stage, flags, environment, the day and the project HEAD. It is valid for 30 min, and only while no project or toolkit file is newer than the run. Any doubt re-evaluates. MXTK_GATE_CACHE_TTL_MIN=0 turns it off.
2. A one-stage query skips the obligation and artifact rows of later stages. In their place it prints one "N row(s) … not evaluated" line, and the summary says so.
3. PROJECT.md is parsed once per run, instead of one awk per lookup.
4. The toolkit is fetched at most once per 15 min (MXTK_FETCH_TTL_MIN).

Evidence

Copy of the purchase-requests base project, Stage 3:

Run	execve before	execve after	Time before	Time after
gate-check	560	275	2.1 s	1.4 s
closeout, cache miss	820	392	3.6 s	2.6 s
closeout, cache hit	820	127	3.6 s	1.5 s
no-stage run	626	338	2.7 s	2.1 s

Closeout and no-stage output are byte-identical, after masking dates and paths. Stage verdict lines and exit codes are unchanged.

Fixtures: all pass.

Fixture	Result
test-bug03-gates	55/0
test-source-ledger	55/0
test-closeout	34/0
test-wrong-verdicts	20/0
test-obligation-*	pass
test-bug02, 04, 05, 06	pass
test-change-log	pass
test-stage-header	pass
test-stage-p	pass
test-surface-column	pass
test-wireframe-count	pass
test-open-questions	43/43

test-bug12-sync fails identically on master.

Risk

• A one-stage query's output changes for rows owed by later stages: they are no longer listed one by one.
• The cache can miss an edit made, during the 1–2 s run, to a file the gate writes itself.
• A "toolkit updated" notice can arrive up to 15 min late.
• Not run on Windows or macOS. The touch -t and date calls have GNU/BSD fallbacks.

---

Generated by Claude Code