[MendixMau]

Direct lane

Lands after #498. This branch was cut from #498's head (2a0886b), so until #498 merges the diff here also shows #498's commits. This PR's own changes are commits 8431714 (the instrument) and 7f91347 (row-21 trigger). #498's head 292ded7 is merged in (a merge, not a rebase). Review this PR's own diff with git diff 292ded7..1858281.

What changed and why (one paragraph):
A script drafted from a DESCRIBE taken before an earlier script of the batch ran put four old messages back, with check, mxbuild and lint green and exit 0 (#261). New project-bin/describe-baseline.sh stamp <script> writes a header from a DESCRIBE at draft time:

• one -- baseline: <Doc> sha256=<hash> line (or absent) per document the script restates;
• one -- baseline-mxcli: <version> line.

exec.sh runs verify only for a script that carries such a header. It prints WARN: stale draft: lines, and DRIFT also goes into the BUILD-LOG detail. It never refuses, and the exit codes (0/1/2/4/5) do not change. verify reports:

• DRIFT, with the first lines of the DESCRIBE diff;
• STALEHEADER when the header names a document the script no longer restates. This is now a warn (exit 1), so exec.sh prints it;
• RESTAMP after an mxcli upgrade: the version recorded at stamp time differs, so a moved hash is not called drift.

A re-run by the script itself (it is the last recorded writer in #498's ledger) and a fresh model (document absent) read SAME.

Preflight STOP 21: its trigger now names create or modify, and the rule gains one clause. There is no new row. It adds the -- supersedes: header, describe-baseline.sh stamp when drafting ahead, and that exec.sh only warns. The DESCRIBE-visibility row (old STOP 34) is dropped.

Fixes #261
Supersedes #311 (with #498)

Behaviour change to check in your project: exec.sh output can carry WARN: stale draft: … lines, and the BUILD-LOG detail cell WARN stale baseline: <docs>. Set MXTK_RESTATE_GUARD=off to skip the check. Scripts stamped by this tool carry -- baseline: / -- baseline-mxcli: header comments.

Impact

Row	Answer
Process	Nothing new blocks. One optional draft-time command (bin/describe-baseline.sh stamp <script>, named in STOP 21).
Dependencies	Touches project-bin/exec.sh (one block after #498's co-owner block) and bin/lib/install-manifest.sh (new installed script). New artifact read: the -- baseline: header. Its producer is stamp, which is opt-in; when there is no header, exec.sh does nothing. It reads #498's ledger through script-owners.sh last-writer, so it depends on #498 and must merge after it.
Context cost	skills/learned-mdl-preflight.md (baseline-routed): +24 words in row 21 (+21 for the rule clause, +3 for the trigger), no new row. render-routing.sh --check: base Baseline: 73804 words (budget 80000) (master 9f3948c reads the same) → head Baseline: 73828 words (budget 80000). Stage 5: mdl 30345 → 30369; nothing else changes.
Behaviour	NOT RUN: the A/B for the changed STOP 21 text (always-loaded) is the maintainer's to run. No agent stub changed.
Runtime	Stub mxcli and mxbuild, wall time, 5 runs each, on an exec of a script with no header. 20 scripts: master 88e9848 1406-1458 ms; #498 1495-1523 ms; this PR 1458-1481 ms. 600 scripts: master 1382-1443 ms; #498 1485-1525 ms; this PR 1482-1584 ms. A script with no header gets one extra grep and zero DESCRIBE calls (fixture-asserted). A stamped script pays one DESCRIBE per baselined document plus one mxcli --version; one DESCRIBE took 0.22 s with real mxcli v0.26.0.

Field evidence: no client project was run. Before the split, exec.sh ran with real mxcli v0.26.0 on a scratch copy of mxcli's public testdata/pedapp model:

1. Script 01 wrote 'OLD wording'.
2. Script 03 was stamped from a DESCRIBE.
3. Script 02 wrote 'NEW wording'.
4. Script 03 then exited 0 and printed WARN: stale draft … Administration.ACT_Probe — changed since the script was drafted, with the real diff (- log node 'Probe' 'OLD wording'; + log node 'Probe' 'NEW wording';).

The golden captures in tests/wave2/fixtures/restatement-guard/ come from that model: the not-found output of all four kinds (shipped in #498) and the mxcli --version line (shipped here). A field run on a real project is still owed.

Fails on base, passes on head: tests/wave2/test-describe-baseline.sh

Subject	Passed	Failed
head 1858281	33	0
#498 head (no describe-baseline.sh)	10	24
#498's exec.sh with the new describe-baseline.sh beside it	30	3
this PR's exec.sh with the earlier 8ad0d77 describe-baseline.sh	28	5

• The 3 failures with fix(exec): warn when replaying an older owning script reverts newer work (#269) #498's exec.sh are every exec.sh case: the WARN line, the drifted document, the BUILD-LOG cell.
• The 5 failures with the 8ad0d77 script are all F8 cases: the version header, re-stamp, RESTAMP (2 assertions), and STALEHEADER as a warn.
• fix(exec): warn when replaying an older owning script reverts newer work (#269) #498's fixture test-restatement-guard.sh still passes 51/0 on this head.

Re-review (head 1858281): test-describe-baseline 33/0, test-restatement-guard 58/0 (8431714 fails the 2 new fresh-clone cases), test-exec-gate-retry 47/0, test-exec-gate-log 24/0, test-restore-mpr-arg 10/0. check-scripts, check-portability, the leak guard, check-pr-discipline: all clean. The client grep is 0. On 600 scripts: master 9f3948c 1405-1440 ms, this head 1480-1529 ms, 0 DESCRIBE calls.

Scoped checks (head 8431714):

• Fixtures:

  • test-describe-baseline 33/0
  • test-restatement-guard 51/0
  • test-exec-gate-retry 47/0
  • test-exec-gate-log 24/0
  • test-exec-check-refusal 10/0
  • test-exec-crash-rollback 13/0
  • test-exec-lock-watch 84/0
  • test-exec-step-times 7/0
  • test-bug07-08 45/0
  • test-manifest-clean-checkout 1/0

• Guards, all clean: check-scripts, check-portability, render-routing --check, the leak guard, check-no-private-citations, check-pr-discipline. The client-name grep of added lines returns 0.

• No client data anywhere in the diff

• Size cap: this PR's own commit touches 7 files with 226 non-fixture lines (+223 −3).

• Test tier reached: T1/T2. A stub-mxcli fixture, plus real mxcli v0.26.0 on a public model. No T3 field project.

• For a new/changed instrument:

  • Golden input is captured (not-found output, --version), never hand-written.
  • Both layouts: the script key and ledger resolve through PROJECT_ROOT (the two-tree case is covered by fix(exec): warn when replaying an older owning script reverts newer work (#269) #498's fixture).
  • Both platforms: bash 3.2 only. Checked by inspection; not run on macOS or Git Bash.

• For a new skill: n/a

• CHANGELOG.md line, credited to AlvinnYeo ([bug] exec.sh: restating a microflow from a stale DESCRIBE silently reverts an earlier exec #261)

• New bug entries: none

---

Generated by Claude Code

[MendixMau]

Behaviour A/B for the row-21 change (RUNBOOK decision rule: T no worse than M on every scored metric)

• M = master c2b9bec, T = this head 1858281. The packs differ only in preflight row 21 (+24 words). Fresh sonnet agent per rep; 66 reps: 10 × mf, 8 × page, 3 × each of the 5 probes, per arm. read_all = 1 in all 66.
• The planted STOP 21 case (mf: whole-document restatement of a shared microflow) is flagged in 10/10 reps in both arms. All 10 T reps now also name -- supersedes: and describe-baseline.sh stamp (0/10 in M), so the clause is read and used as intended.
• Quoting (mf quoted): on first scoring, two T reps came out at 55% and 61%. The cause: the T agents follow the new clause and write header comments such as -- supersedes: and DESCRIBE MICROFLOW Sales.X, and score.sh counts unquoted names inside comments. Counting only the MDL code lines, M and T are the same: median 91% in both, 8 of 10 reps at 91% or higher in both.
• The other metrics are equal or differ by one rep: probes probe_first the same in both arms; sections, report, planted, vfb, cinloop, rinloop, pos and txtcrumb identical; page crumbs 3/3 in 5 of 8 reps in both arms (T has one 0/3, M none below 2/3); mf folder complete in 7 vs 6 reps; one T page rep wrote an extra H1.
• Verdict: T is no worse on behaviour. Follow-up for the scorer: quoted should skip comment lines.

---

Generated by Claude Code