fix(exec.sh): mxbuild delta gate keeps a write that reduces the pre-existing error set

CHANGELOG.md

@@ -84,6 +84,7 @@ Sections dated before 2026-09-19 predate the cycle and stay as they are.
 - learn(skills/lint-that-actually-runs.md): **three more lint vocabulary facts, verified on v0.24.0, with a before/after rule.** `microflow_type` is `MICROFLOW`/`NANOFLOW`/`RULE`, and `microflows()` yields all three. SQL `PERSISTENT` is Starlark `Persistent`. `activity_count` counts top-level objects (splits, loops, annotations), not actions, and `activities_for()` does not see loop bodies. — field report from a colleague's naming-conventions project (mxcli v0.23.0, Mendix 11.12.4)
 - learn(skills/learned-microflow-patterns.md): **expression functions take positional arguments only — `toString(from: $X)` is not a syntax error, it's a silent misparse.** MDL's expression grammar has no named-argument form for function calls; `label: value` parses as `label` (a bare unresolved identifier) COLON-divided by `value` — COLON being the OQL division operator. Verified on mxcli v0.24.0 / Mendix 11.12.1, scratch copy of a PoC model: `toString(from: $Temperature)` and `formatDateTime($D, pattern: 'yyyy-MM-dd')` both pass `mxcli check --references` silently (exit 0) because the collapsed argument count still matches the function's arity; `substring(from: $S, index: 0, length: 3)` fails, but with a misleading `E006` ("expects 2 to 3 argument(s), got 1") that never names labelled arguments as the cause. The likely source: labelled-colon syntax is genuine MDL elsewhere (page `Attribute:`/`Action: MICROFLOW Name(Param: val)`, `@anchor(from: bottom, to: top)`) — just never inside an expression — workshop feedback, Sep 2026
 - fix(bin/doctor.sh): **the gate self-test's scratch copy now carries the whole model directory, not just `.mpr` + `mprcontents/`.** A `--target=deploy` build (the same target the real gate runs, in place, against the project as it sits on disk) resolves widget/theme/design-property references out of `theme/`, `resources/`, `widgets/` and `javasource/` sitting beside the `.mpr`; a thin copy reported hundreds to over a thousand `Could not find widget` / design-property errors as a "dirty baseline" while the real gate passed with 0 errors on the identical model. Now copies the directory holding the `.mpr` (single-tree root or `app/` on a two-tree checkout) minus `.git/`, `deployment/`, `node_modules/` and `.mpr-snapshots/`, and prints the copied size (`du -sh`) so the cost is visible. Keeps the existing real-basename fix. A sibling `.mpr` in the project's own directory was ruled out (collides with the project's own `mprcontents/`); `mxbuild --target=check` was ruled out (no evidence it exists anywhere in this toolkit's usage) — field reports, issue #127 and a macOS custom-theme project
+- fix(project-bin/exec.sh): **the mxbuild delta gate now keeps a write when the post-exec error set is a strict subset of the pre-flight baseline, not just when it is identical.** A script that took 34 pre-existing errors down to 1 was being rolled back and logged `blocked: PRE-EXISTING CE1613`, because the gate could only recognise "unchanged," never "reduced." The comparison key is unchanged (`err_set`'s per-message string); a new `is_subset_of` helper checks membership, and a set containing anything NEW still restores the snapshot exactly as before. The kept-write BUILD-LOG row now reads `⚠️ applied (dirty model, reduced)` with detail `errors 34→1 (pre-existing, reduced); remaining [...] still blocks deploy` — field report from a colleague's naming-conventions project (mxcli v0.23.0)
 - learn(skills/learned-file-upload-widget.md): **a file upload mxcli can author, with proof that it uploads.** The Mendix File Uploader 2.5.0 bound to a `System.FileDocument` specialisation: the MDL shape (entities, grants, create/delete microflows, advanced formats), which upload widgets mxcli cannot author (classic FileManager, PDS uploader), the two traps with workarounds (a simple-mode `allowedfileformat` passes exec and fails `mx check` with CE0463; an uploader DESCRIBE will not re-exec, `exposes 2 datasources`), and the six-step upload instrument. Field run on stock v0.24.0: the section-4 MDL taken verbatim from the skill gave `mx check` 0 errors, 2/2 files stored, 2/2 downloads sha256-equal, `.csv` rejected with 0 rows; both traps reproduce unchanged on v0.24.0 — a Mendix app-rebuild project
 - new(skills/mendix-best-practices-index.md): **one row per Mendix best-practice area: the Mendix docs page, the bundled `assess-quality` section, the toolkit skill that applies it before the write, and the `mxcli lint` rule that catches it after exec.** An index, not a copy — the practice text stays on the Mendix pages (17 URLs verified HTTP 200 on 2026-09-25) and in the mxcli-bundled skill; Mendix's own Best Practice Recommender rules (MXP001–016) anchor the performance rows, and four rows say out loud that no lint rule exists and the preflight checklist is the only check. Routed `all` agents, stages 3/5/6, group reference — Maurits Visser
 - learn(skills/microflow-preflight.md, agents/mdl-agent.md): **a microflow now gets a tier before any MDL — Simple, Guided or Split-first — and Split-first means a posted split plan (thin orchestrator + one `SUB_` per responsibility, with signatures) that the user confirms first.** Prompted by a colleague's session refusing a long microflow as "too difficult" while the same task went through on a stronger model: the piece was too big, not the task. mdl-agent gains two rules — never hand back "too difficult", hand back the split plan; escalate one failing `SUB_` by name after one retry, never the whole script. Also records the mxcli team's answer on positioning: a standalone `mxcli layout` command — which on v0.24.0 and upstream main (2026-09-25) arranges domain models only (`--dry-run` on a scratch copy of a PoC model: 10 entity moves, no microflow), so the no-`@position` rule and the if-branch workaround stand until a microflow mode ships; noted in the bug ledger and `learned-microflow-patterns.md` — Maurits Visser

README.md

@@ -313,7 +313,7 @@ Use this before every write. Full per-rule detail (root causes, bug IDs, retest
 | Drop an attribute that has security grants | Studio Pro GUI | N/A |
 | After any MPR corruption or load error | `bin/restore-mpr.sh` | Closed |
 
-**The crash net.** An MPR is two parts: `Project.mpr` (SQLite index) and `mprcontents/` (BSON units). `bin/exec.sh` snapshots both before every batch; 5 rotate; `bin/restore-mpr.sh` rolls back both together (either alone is useless). Git commits at phase gates are the real history. Ad-hoc `.mpr.backup` copies are banned.
+**The crash net.** An MPR is two parts: `Project.mpr` (SQLite index) and `mprcontents/` (BSON units). `bin/exec.sh` snapshots both before every batch; 5 rotate; `bin/restore-mpr.sh` rolls back both together (either alone is useless). Git commits at phase gates are the real history. Ad-hoc `.mpr.backup` copies are banned. By design, `exec.sh` refuses to run at all while the model has uncommitted changes — its snapshot would not cover them, so a later auto-restore could silently lose that work; commit the model first (`FORCE_EXEC=1` overrides, at your own risk).
 
 ### Something went wrong? Don't panic.
 

project-bin/exec.sh

@@ -2,7 +2,15 @@
 # exec.sh — the guard chain around a model write.
 #
 #   concurrent-writer guard → module-brief advisory → mxcli check → snapshot → baseline → exec
-#   → mxbuild gate → auto-restore on regression → lint ratchet → SP reopen
+#   → mxbuild delta gate → auto-restore only on a NEW error → lint ratchet → SP reopen
+#
+# The mxbuild gate is a DELTA gate, not an absolute "0 errors" one: a write is kept
+# when the post-exec error set is the pre-flight baseline exactly, or a strict
+# SUBSET of it (every post-exec error message already existed before this script
+# ran — nothing new, even if some were cleared). Only a set containing something
+# new triggers the snapshot restore. See "Delta gate" below.
+# Known limit: errors are compared by message text only, so a new error whose
+# message matches one already in the baseline is not seen as new.
 #
 # Usage: ./bin/exec.sh <script.mdl>
 #        ./bin/exec.sh --patch <script> [args...]
@@ -23,6 +31,10 @@
 # the project's allow-list (install-claude-permissions.sh), so in auto mode Claude Code's
 # classifier does not review the script. Read what a patch does outside the .mpr before running it.
 #
+# Refuses to run while the model has UNCOMMITTED changes (by design: the snapshot
+# taken below would not cover them, so a later auto-restore could silently lose
+# that work) — commit the model first, or FORCE_EXEC=1 to override at that risk.
+#
 # Overrides: FORCE_EXEC=1 (skip refusals), SKIP_CHECK=1 (skip the pre-exec
 #            mxcli check), SKIP_BASELINE=1 (skip pre-flight mxbuild),
 #            MXBUILD_PATH=..., MENDIX_APP=..., MPR_FILE=...,
@@ -547,6 +559,34 @@ err_set() {
   "$PY" -c "import json;d=json.load(open('$(native_path "$1")'));print('|'.join(sorted(x.get('message','') for x in d.get('problems',[]) if x.get('severity')=='Error')))" 2>/dev/null || echo ""
 }
 
+# is_subset_of <candidate> <superset> — both are err_set's own "|"-joined sorted
+# message strings (the SAME comparison key the identical-baseline check already
+# used; this does not introduce a new one). True (exit 0) when every message in
+# <candidate> also appears in <superset> — i.e. nothing NEW. An empty <candidate>
+# (0 post-exec errors) is trivially a subset. Bash-native, no extra Python call:
+# both strings are already in hand as plain variables by the time this runs.
+# Pure POSIX word-splitting on IFS='|', no arrays/`read -a` — bash 3.2 / Git Bash
+# safe. Messages containing a literal "|" would break the membership test the
+# same way they already break the exact-equality test above; not new exposure.
+is_subset_of() {
+  _cand="$1"
+  [ -z "$_cand" ] && return 0
+  _super="|$2|"
+  _oldIFS="$IFS"
+  IFS='|'
+  for _e in $_cand; do
+    IFS="$_oldIFS"
+    [ -z "$_e" ] && continue
+    case "$_super" in
+      *"|$_e|"*) : ;;
+      *) return 1 ;;
+    esac
+    IFS='|'
+  done
+  IFS="$_oldIFS"
+  return 0
+}
+
 # ── Doctor freshness (warn-only) ─────────────────────────────────────────────
 # doctor.sh used to run once, at install, and never again. A machine that drifts mid-project
 # (new mxcli, a wrong-arch mxbuild, an endpoint agent slowing every fork) then reads as "the
@@ -724,14 +764,39 @@ if [ -x "$MXBUILD" ] && [ -x "$JAVA_EXE" ]; then
       # ── Delta gate ─────────────────────────────────────────────────────────
       # A shared model means another workstream can leave it non-building; an
       # absolute "zero errors" gate would then block every good script forever.
+      # Keeps the write when the POST-exec error set is the baseline exactly
+      # (nothing changed), OR a STRICT SUBSET of it (every post-exec message
+      # already existed pre-exec — no new message, even though some
+      # pre-existing ones may have been cleared). Messages only: a new error
+      # with the same text as a baseline one passes (known limit, see header).
+      # Only a set with something NEW in it restores the snapshot.
+      # Real incident: a script that took 34
+      # pre-existing errors down to 1 was rolled back and logged
+      # "blocked: PRE-EXISTING CE1613" because the gate could only recognise
+      # "unchanged," never "reduced." Comparison key is unchanged — err_set's
+      # per-message string (see err_set above) — subset-checked by
+      # is_subset_of, not switched to a different key.
       POST_SET=$(err_set "$ERRORS_FILE")
+      DELTA_KIND=""
       if [ -n "$BASELINE_SET" ] && [ "$BASELINE_SET" = "$POST_SET" ]; then
+        DELTA_KIND="identical"
+      elif [ -n "$BASELINE_SET" ] && is_subset_of "$POST_SET" "$BASELINE_SET"; then
+        DELTA_KIND="subset"
+      fi
+      if [ -n "$DELTA_KIND" ]; then
         GATE_STATE="pass"
         KEEP_CODES=$(err_codes "$ERRORS_FILE")
-        echo "  ⚠  mxbuild: $CE_COUNT error(s) — IDENTICAL to the pre-flight baseline."
-        echo "     Pre-existing [$KEEP_CODES], not introduced by this script. KEEPING the changes."
-        echo "     The model still will not deploy until those are cleared in Studio Pro."
-        [ "$EXEC_STATUS" -eq 0 ] && log_build "⚠️ applied (dirty model)" "no new errors; pre-existing $KEEP_CODES still blocks deploy"
+        if [ "$DELTA_KIND" = "subset" ]; then
+          echo "  ⚠  mxbuild: $CE_COUNT error(s) — a STRICT SUBSET of the $_BC pre-flight baseline error(s), no new ones."
+          echo "     Pre-existing [$KEEP_CODES] remain; this script reduced the error count. KEEPING the changes."
+          echo "     The model still will not deploy until those are cleared in Studio Pro."
+          [ "$EXEC_STATUS" -eq 0 ] && log_build "⚠️ applied (dirty model, reduced)" "errors ${_BC}→${CE_COUNT} (pre-existing, reduced); remaining [$KEEP_CODES] still blocks deploy"
+        else
+          echo "  ⚠  mxbuild: $CE_COUNT error(s) — IDENTICAL to the pre-flight baseline."
+          echo "     Pre-existing [$KEEP_CODES], not introduced by this script. KEEPING the changes."
+          echo "     The model still will not deploy until those are cleared in Studio Pro."
+          [ "$EXEC_STATUS" -eq 0 ] && log_build "⚠️ applied (dirty model)" "no new errors; pre-existing $KEEP_CODES still blocks deploy"
+        fi
         cp "$ERRORS_FILE" "$LAST_ERRS"
         rm -f "$ERRORS_FILE"
       else