Integration 2026-10-20: six reviewed bug fixes (#486 #480 #482 #481 #484 #485)

CHANGELOG.md

@@ -20,8 +20,14 @@ Sections dated before 2026-09-19 predate the cycle and stay as they are. Named m
 ## Unreleased
 
 - fix(gates): **a decision taken during a stage no longer reads as that stage's approval** (#477). The #207 flag prefixed a FAIL with `APPROVED OVER A FAILING GATE` whenever any CONFIRMED row carried the stage's number, so a Stage-5 build with `| 5 | Sub-agents run on Sonnet | CONFIRMED |` was flagged, and `status.sh` told the lead to re-open an approval nobody gave while the real gap (PENDING obligations) sat underneath. The register has no dedicated exit row, so `gate-check.sh` now counts a stage as approved on either of two pieces of evidence in the Decisions table: an explicit exit row ("Stage N approved", "Stage 4 — approved with conditions", "Approved: Stage N …", "Gate approved", or "<artifact> approved" for a fixed artifact list such as "Build plan approved"), or the stage having CONFIRMED rows of its own plus a CONFIRMED decision for a later stage up to 6 (the #207 field case); an early Stage-7 "no cutover" row or an N/A/skip row is not evidence. The Stage 3/4 gate checks and every verdict and exit code are unchanged. `test-bug03-gates.sh` T17 (16 cases, checked under gawk, mawk and original-awk): 7 FAIL on master, 47/47 with the fix. — MendixMau
+- fix(progress-report): **the top bar is one group of section links (Stages, Build, Pages, Tips) instead of stage chips plus links; the page carries one `id=stamp`, and the header drops the build-plan.html note.** The sticky bar loses the `Stage P 0 … 6` chips, which repeated the stage list right below it, and gains a `Stages` link to that list; the `#stage-N` card anchors stay for links from other pages, and the header now states the current stage. Five elements shared `id=stamp` (invalid HTML); the "Last updated" span keeps it and the others get their own ids plus a `data-stamp` marker, which `build-plan-status.sh --refresh` now reads (with `id=stamp`, so pages written before this still compare) to leave an unchanged page alone. The filter now depends on both scripts: a project whose `bin/build-plan-status.sh` is locally edited keeps the old `id=stamp` filter after sync, so the page is rewritten on every refresh; such a project needs `sync-project.sh --only-bin build-plan-status.sh --upgrade-bin build-plan-status.sh` alongside this. "replaces the former build-plan.html" is gone from the header sub-line. Fixes #479 items 1–3 — MendixMau (review 2026-10-20)
+- fix(conformance): **a definition that merely contains "(0 " or "not found" is no longer scored ABSENT** (#470). `conformance-check.sh` matched both markers anywhere in mxcli's answer, so a built microflow whose documentation read "HttpStatus (0 = no response or time-out)" went STALE (five false STALEs in one module), and so does every `describe navigation` with a not-found page, which mxcli prints as "  not found page <Page>". It now reads an error only on the first non-blank line (`Error:…`, the old start-of-output meaning; enumeration, scheduled-event and image-collection docs print raw, so a doc line may start `Error:`), and an empty set only when the last non-blank line is exactly the `(0 …)` summary (also `(0 workflow group(s))`); CRLF output reads the same. New fixture `tests/wave2/test-conformance-absent-marker.sh` replays two verbatim captures already in the repo (the v0.25.0 navigation describe of mxcli's public test app, and a v0.23.0 microflow describe with the issue's doc sentence spliced in), plus an enumeration with a raw `Error:` doc line and CRLF-only cases, both layouts; 6 FAIL on master, 24/24 with the fix. — jopterhorst
+- fix(page-fidelity): **a popup wireframe is scored against its dialog, not the page drawn behind it** (#471). A popup wireframe draws the caller's page behind a scrim (`.popup-stage > .shell-behind + .dialog-scrim > .dialog`); `page-fidelity.js` looked for `div.main` first and landed in the shell, so a popup whose fields and bindings all matched scored 29% in the field, every miss being the caller's page head, h1 and card. With popup evidence (a `.popup-stage` or `.dialog-scrim` in the wireframe, or a page whose MDL names a popup layout) the scorer now scores the `.dialog` only (class word, so `dialog-head` does not count), or the stage minus `.shell-behind` when there is no `.dialog`. Without that evidence nothing changes, so a full-page wireframe that draws an inline `.dialog` confirm (the design system's own component) scores exactly as before, and a bare-`.dialog` popup wireframe still reads `<body>`. `test-page-fidelity-mocks.sh` gains the popup cases on a verbatim v0.23.0 describe of a real popup page and a wireframe drawn in the issue's shape, plus layout-only evidence and the full-page-plus-inline-confirm regression case: 18% and 8 FAIL on master, 100% and 78/78 with the fix. — jopterhorst
+- fix(monkey): **the monkey no longer signs itself out of a non-English app** (#469). `monkey.js` recognised only English sign-out captions, so in a Dutch app it clicked "Afmelden", landed on `login.html`, scored a false CRASH and fuzzed nothing more (seed 204262061: 3 observations, 1 crash-class; 24 and 0 with the label excluded). The default now also covers Dutch, German, French, Spanish, Portuguese and Italian sign-out captions, matched as the whole caption so ordinary verbs stay fuzzable ("Ziek afmelden", "Vom Kurs abmelden", and sign-in captions such as "Aanmelden"); a project adds its own in `project.config.js` `signOutPattern` or `SIGN_OUT_PATTERN`, added to the default, never replacing it, with its own regex flags kept. `rapidTab` no longer presses Enter while focus is on a sign-out control (top bar, nav item, aria-label or title, each tested on its own), the route by which a layout control outside `.mx-page` was reached. `test-monkey-left-app.sh` gains the issue's captions, sign-in near-misses, the over-match captions, the project pattern and its flags, and a rapidTab guard scoped to that branch (it fails on a blind-Enter mutation): 9 FAIL on master, 22/22 with the fix. — jopterhorst
 - fix(progress-report): **the build plan is shown once: the phase cards are the plan.** Expanding a card now shows its full rows (#, Kind, Step, Produces/Proves, Depends on, State, Scripts) and Claims, which only the duplicate "Full plan tables" block carried; that block is gone and views A (script folders) and B (module proof) are their own section. The `now` phase renders open with a selected style, opening another phase closes the previous one (accordion; `#phase-N` links still open their card), and a viewer's saved open/closed state survives the 30 s reload and wins over the default (#474) — field feedback from a requirements-driven project at Stage 5 — MendixMau (field feedback 2026-10-20)
 
+- fix(status): **`status.sh --brief` no longer aborts with "LAST_FAIL<byte>: unbound variable" before printing NEXT (#467).** Two NEXT lines wrote a bare `$LAST_FAIL`/`$NOT_DONE_PASS` straight before an em-dash; in a single-byte locale (and on macOS bash 3.2 even under `en_US.UTF-8`) bash reads the dash's first byte as part of the name and `set -u` kills the run. Braced there and at the two other sites with the same idiom (`project-bin/constants-audit.sh`'s `…`, a case pattern in `test-runbook-slice.sh`). `check-portability.sh` now flags any bare `$NAME` followed by a non-ASCII byte, in every file including the meta-exempt ones (matched with `LC_ALL=C`, so the check does not depend on the locale it guards against; `# portability-ok:` opts out). New `tests/wave2/test-status-bare-var.sh`: the lint cases on any machine, plus both NEXT rules run under an ISO-8859-1 locale where one exists (reproduced on bash 5.2 with a compiled `en_US.ISO-8859-1`: both rules aborted on master, both print after). — jopterhorst
+- fix(portability): **`[ \t]` in a sed or grep bracket no longer trims a trailing "t" (#468).** BSD/macOS sed reads `[ \t]` as "space, backslash or t" (only GNU sed expands `\t` there, and grep never does), so `context-pack.sh` printed `Mod.ENUM_Product` as `Mod.ENUM_Produc` and called it "not in the folder plan"; `closeout.sh`'s cell trim cut "bulk import" to "bulk impor" and "test" to "est"; `entry-mode.sh` read an answer written `answered:<TAB>migration` as no entry mode at all; and `gate-check.sh`'s intake section count matched `##t…` but never `##<TAB>`. All five sites now use `[[:blank:]]`, and a test fixture's `[^\t]` uses a real tab (`$'…'`). `check-portability.sh` flags `\t` inside a bracket that follows `sed`/`grep` on the line (awk expands `\t` everywhere and is not flagged; a `$'…'` string is skipped). Reproduced on GNU sed through a `sed --posix` shim, which reads the bracket the BSD way: the captured context-pack golden brief (`Sales.Product`, `Sales.ACT_ApprovalStep_Reject`) failed 4 of 4 new BSD-sed checks on master and passes after; new `tests/wave2/test-sed-tab-bracket.sh` covers `closeout.sh`, `entry-mode.sh` and the lint (6 of 9 fail on master). — jopterhorst
 ## v2026.10.19
 
 - new(progress-report): **the progress report gains a Project pages section: a compact folder tree of the HTML pages found in the project's output folders (gate dashboard, analysis reports, blueprint, design system, process map, wireframes, UI reviews, test reports), each a relative link from `architecture/` with its modified date.** Pages are found by scanning the project's files on every render, never from a stored list; only the top-level folders in `PAGE_DIRS` are scanned (each named with its producer; a symlinked one is skipped), `node_modules`, `.mpr-snapshots`, `deployment`, `mprcontents`, `theme`, `.git`, vendor and dot folders are never entered, a folder with more than 8 pages folds behind a count, every folder keeps its row and lists at most 100 links (`+K more`), and the report never lists itself. Dates are shown as days, so the section is unchanged on the same day unless a listed page changed and a `--refresh` leaves the page alone. Test: `tests/wave2/test-progress-report-pages.sh` (single and two-tree layouts, Windows hrefs) — MendixMau (request 2026-10-19)

bin/check-portability.sh

@@ -67,6 +67,21 @@ while IFS= read -r f; do
     report "$f" 1 "CRLF line endings" "bash fails on line 1. Ensure .gitattributes has *.sh text eol=lf."
   fi
 
+  # 2. A bare $NAME followed directly by a non-ASCII byte (an em-dash, an ellipsis). In a
+  # single-byte locale bash reads that byte as part of the variable NAME (macOS bash 3.2 did
+  # it under en_US.UTF-8 too), so set -u aborts with "NAME<byte>: unbound variable" (#467:
+  # status.sh --brief died before printing NEXT). Checked before is_meta: no file is exempt
+  # from this one. Matched byte-wise (LC_ALL=C) so the result does not depend on the locale
+  # it is meant to guard against. Fix: brace it, ${NAME}—.
+  while IFS= read -r hit; do
+    [ -n "$hit" ] || continue
+    case "$hit" in *portability-ok:*) continue ;; esac
+    report "$f" "${hit%%:*}" "bare \$NAME followed by a non-ASCII character" \
+      'Brace the name: ${NAME}. A single-byte locale (and bash 3.2) reads the first byte as part of the name; set -u then aborts.'
+  done <<EOF
+$(code_lines "$f" | LC_ALL=C grep -E '^[0-9]+:.*\$[A-Za-z_][A-Za-z0-9_]*[^[:print:][:cntrl:]]')
+EOF
+
   is_meta "$f" && continue
 
   while IFS= read -r hit; do
@@ -157,6 +172,21 @@ while IFS= read -r f; do
         fi ;;
     esac
 
+    # \t inside a sed/grep bracket expression is not a tab. BSD/macOS sed reads [ \t] as
+    # "space, backslash or t" (GNU sed alone expands it), and grep reads it that way on
+    # every platform, so a trim stripped a trailing "t": Mod.ENUM_Product became
+    # Mod.ENUM_Produc in context-pack.sh (#468). awk expands \t in a regex everywhere, so
+    # only a bracket that FOLLOWS sed/grep on the line counts; a bracket inside a $'...'
+    # string holds a real tab and is skipped.
+    case "$txt" in
+      *sed*'\t'*|*grep*'\t'*)
+        if grep -Eq -- '(^|[^[:alnum:]_])(sed|grep)([[:space:]]|$)[^|]*\[[^]]*\\t' <<<"$txt" \
+           && ! grep -Eq -- "\\\$'[^']*\\[[^]]*\\\\t" <<<"$txt"; then
+          report "$f" "$ln" "\\t inside a sed/grep bracket expression (BSD: backslash or t, not tab)" \
+            'Use [[:blank:]] (space or tab) or [[:space:]], or a real tab via $'"'"'\t'"'"'.'
+        fi ;;
+    esac
+
     case "$txt" in
       *'grep -P'*|*'grep -qP'*|*'grep -oP'*)
         report "$f" "$ln" "grep -P is GNU-only" \

bin/gate-check.sh

@@ -761,7 +761,7 @@ check_stage_P() {
       else if (l ~ /^unverified/ && l ~ /how to verify/) ok=1
     }
     END { flush(); print bad }' "$f")
-  total=$(grep -c '^##[ \t]' "$f")
+  total=$(grep -c '^##[[:blank:]]' "$f")
 
   # Always name the file that was graded. A Stage P line that does not is what let the
   # wrong-file bug survive for weeks: the output was locally coherent and globally wrong.

bin/lib/closeout.sh

@@ -107,7 +107,7 @@ _co_checkpoints() {
   esac
 }
 
-_co_trim() { printf '%s' "$1" | sed -E 's/^[ \t]+//; s/[ \t]+$//; s/^\*+//; s/\*+$//; s/^[ \t]+//; s/[ \t]+$//'; }
+_co_trim() { printf '%s' "$1" | sed -E 's/^[[:blank:]]+//; s/[[:blank:]]+$//; s/^\*+//; s/\*+$//; s/^[[:blank:]]+//; s/[[:blank:]]+$//'; }
 # Truncate to N chars on a word boundary, with an ellipsis. Cells in a real register run to
 # 2,000 characters; the block is a recap, the register is the record.
 _co_trunc() {
@@ -117,7 +117,7 @@ _co_trunc() {
 }
 # Stage cell → bare stage token ("Stage 3" → 3, "P" → P). Empty when it is not a stage.
 _co_stage_of() {
-  local s; s="$(_co_trim "$1" | sed -E 's/^[Ss]tage[ \t]*//')"
+  local s; s="$(_co_trim "$1" | sed -E 's/^[Ss]tage[[:blank:]]*//')"
   case "$s" in P|p) echo P ;; [0-7]) echo "$s" ;; *) echo "" ;; esac
 }
 # First word of a status cell, upper-cased, punctuation stripped — the same "controlled

bin/lib/entry-mode.sh

@@ -34,14 +34,14 @@ entry_mode_token() {
   # doing it after would let that stripping treat "(2026-09-01)" as if it opened at the
   # start of the value and eat everything from "assumed (" onward, value included. The
   # register's bare value ("migration") has no such prefix and passes through untouched.
-  v="$(printf '%s' "$v" | sed -E 's/^[ \t>*_-]*(answered|answer|a|decision|assumed)[ \t]*(\([^)]*\))?[ \t]*:[ \t]*//')"
+  v="$(printf '%s' "$v" | sed -E 's/^[[:blank:]>*_-]*(answered|answer|a|decision|assumed)[[:blank:]]*(\([^)]*\))?[[:blank:]]*:[[:blank:]]*//')"
   v="${v%%(*}"          # drop a trailing parenthetical: "migration (not greenfield)"
   v="${v%%,*}"          # drop anything after the first comma
   v="${v%%;*}"          # ...or semicolon
   # Trim leading/trailing noise, INCLUDING markdown emphasis markers on both ends
   # ("**migration**", "_greenfield_") — the old trailing trim only stripped spaces,
   # tabs and a period, so a bolded or italicised value never matched anything.
-  v="$(printf '%s' "$v" | sed 's/^[ \t>*_-]*//; s/[ \t.*_-]*$//')"
+  v="$(printf '%s' "$v" | sed 's/^[[:blank:]>*_-]*//; s/[[:blank:].*_-]*$//')"
   full="$v"
   # Documented multi-word phrases resolve on the WHOLE value, before the first-word
   # truncation below: the runbook's own label for the fourth mode is "Change an existing

bin/status.sh

@@ -240,9 +240,9 @@ next_action() {
   esac
   if [ "$UNSYNCED" -gt 0 ]; then echo "ba-agent flushes $UNSYNCED UNSYNCED marker(s) in PROJECT.md — every gate is blocked until then (conversion-runbook.md §3b)"; return; fi
   case "$COH" in *"DUE"*) echo "run process-coherence-pass.md — the cross-module seam check is due"; return ;; esac   # process-coherence-pass.md
-  if [ -n "$LAST_FAIL" ]; then echo "last gate FAILED for: $LAST_FAIL— fix and re-exec (docs/BUILD-LOG.md, .mpr-snapshots/last-mxbuild-errors.json)"; return; fi   # iterative-build-loop.md Gate: BUILD
+  if [ -n "$LAST_FAIL" ]; then echo "last gate FAILED for: ${LAST_FAIL}— fix and re-exec (docs/BUILD-LOG.md, .mpr-snapshots/last-mxbuild-errors.json)"; return; fi   # iterative-build-loop.md Gate: BUILD
   if [ -n "$DONE_STALL" ]; then echo "$DONE_STALL — done- needs the coverage checklist; run bin/coverage-preflight.sh to see what blocks it (iterative-build-loop.md)"; return; fi
-  if [ -n "$NOT_DONE_PASS" ]; then echo "gate-passed but not done-: $NOT_DONE_PASS— walk the happy path + coverage checklist, then git mv to done- (iterative-build-loop.md step 13-15)"; return; fi
+  if [ -n "$NOT_DONE_PASS" ]; then echo "gate-passed but not done-: ${NOT_DONE_PASS}— walk the happy path + coverage checklist, then git mv to done- (iterative-build-loop.md step 13-15)"; return; fi
   if [ -n "$OB_NEXT" ]; then echo "discharge: $OB_NEXT (bin/lib/obligations.tsv names the artifact each owes)"; return; fi
   [ -n "$GC_NEXT" ] && { echo "$GC_NEXT"; return; }
   echo "nothing overdue — take the next script in architecture/build-plan.md"

project-bin/build-plan-status.sh

@@ -739,8 +739,10 @@ if [ "$WRITE_HTML" -eq 1 ]; then
     if "$PY" "$GEN" "$ROOT" "${GEN_ARGS[@]}" >/dev/null 2>"$OUT.err.$$"; then
       # Two renders of the same state differ only in the stamp lines. A refresh that changed
       # nothing leaves the file alone, or every exec.sh run would dirty the working tree.
+      # Stamp lines carry a data-stamp attribute (one of them also id=stamp); 'id=stamp' is kept in
+      # the pattern so a page written before #479, with id=stamp on every such line, still compares.
       if [ "$REFRESH" -eq 1 ] && [ -f "$OUT" ] \
-         && [ "$(grep -v 'id=stamp' "$OUT")" = "$(grep -v 'id=stamp' "$OUT_TMP")" ]; then
+         && [ "$(grep -vE 'id=stamp|data-stamp' "$OUT")" = "$(grep -vE 'id=stamp|data-stamp' "$OUT_TMP")" ]; then
         rm -f "${OUT_TMP:?}"
       else
         mv -f "$OUT_TMP" "$OUT"

project-bin/conformance-check.sh

@@ -360,10 +360,31 @@ probe() {
   esac
   if [ $rc -ne 0 ]; then echo "ABSENT"; return; fi
   if [ -z "$(printf '%s' "$out" | tr -d '[:space:]')" ]; then echo "ABSENT"; return; fi
-  case "$out" in
-    Error:*|*"not found"*) echo "ABSENT"; return ;;
-    *"(0 "*)               echo "ABSENT"; return ;;   # empty result set, e.g. "(0 entities)"
-  esac
+  # Read the two absence markers only where mxcli prints them, never in the definition body.
+  # A substring match read a DESCRIBE whose documentation said "HttpStatus (0 = no response or
+  # time-out)" as an empty set, and the verbatim v0.25.0 `describe navigation` of mxcli's public
+  # test app carries "  not found page Pages.Vehicle_Overview" (a navigation setting): both
+  # scored ABSENT, so built rows went STALE (#470).
+  #   Error   — mxcli puts its error FIRST (stderr is merged above, and `-c` pairs it with exit 1,
+  #             which the rc test already scored); so only the first non-blank line is read, the
+  #             old `Error:*` start-of-output meaning. Not any line: a body line CAN start at
+  #             column 0 — enumerations, scheduled events and image collections print their
+  #             documentation raw inside `/** … */` (one ` * ` on the first line only), and a
+  #             multi-line string prints raw too, so a doc line "Error: …" is still PRESENT.
+  #   (0 …)   — a table's summary is printed as its own LAST line, `fmt.Fprintf("\n%s\n", Summary)`
+  #             with Summary = "(%d entities)", "(%d workflow group(s))", …; a definition ends in
+  #             `/`, `;` or a grant, never in a parenthesised count.
+  # CR stripped for a Windows-built mxcli (printed "\n%s\n", so nothing else trails the summary).
+  # Here-strings, not `printf | grep -q`: under pipefail
+  # that pipe can lose to SIGPIPE on a long definition (test-pipefail-grep-q.sh).
+  local first last
+  out="$(tr -d '\r' <<<"$out")"
+  first="$(sed -n '/[^[:space:]]/{p;q;}' <<<"$out")"
+  case "$first" in Error:*) echo "ABSENT"; return ;; esac
+  last="$(sed '/^[[:space:]]*$/d' <<<"$out" | tail -n 1)"
+  if grep -Eq '^\(0 .*\)$' <<<"$last"; then   # no trailing class: the CR strip above is what lets CRLF output match
+    echo "ABSENT"; return   # empty result set, e.g. "(0 entities)"
+  fi
   echo "PRESENT"
 }
 

project-bin/constants-audit.sh

@@ -167,7 +167,7 @@ while IFS= read -r line; do
   # four sentences — and printing it whole turns a 30-row table into a wall. The register is
   # the place to read it; this column only has to prove a reason exists.
   short=$(printf '%s' "$reason" | cut -c1-58)
-  [ "${#reason}" -gt 58 ] && short="$short…"
+  [ "${#reason}" -gt 58 ] && short="${short}…"
 
   ROWS="$ROWS$(printf '%-13s %-44s %-10s %-8s %s' "$verdict" "$qname" "$type" "$state" "$short")
 "