MendixMau · Nov 1, 2026 · Oct 28, 2026 · Nov 1, 2026 · Nov 1, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -38,6 +38,7 @@ Sections dated before 2026-09-19 predate the cycle and stay as they are. Named m
 - fix(page-audit, design-audit): **the audits produce their own `.claude/loop/page-scope.json`**: when it is missing or older than the model they run `bin/page-scope.sh` (`scope-ensure.js`, 2 s) instead of FAULTing on 0 pages; a failed regeneration never blanks an existing file and names the failure. Field: scope deleted -> page-audit ran on 9 pages, `touch` of the .mpr -> regenerated once — speed experiment E2
 - new(look-ledger, journey-runner, module-review): **a journey run's screenshots are LOOK evidence**: `bin/look-ledger.sh shots` (run by journey-runner after every run) maps each reached step's PNG to the one page whose widgets (`DESCRIBE PAGE`) it landed on, a URL pattern breaking ties, and lists every owed page as `OPEN` (open that PNG; the Read hook then marks the page SEEN) or `UNREACHED` (short gap-fill journey, or `page-audit.js` as before) with the line `owed N = A OPEN + B UNREACHED`; a shot older than the page's last build is not evidence and nothing is claimed when unsure. Field: 8 owed pages = 5 OPEN + 3 UNREACHED from two journey runs, no separate screenshot pass; baseline words 64,004 -> 64,193 — speed experiment E2
 - fix(affected-tests): **a page, flow or navigation change no longer selects 0 journeys because the journeys name widgets, not `Module.Name`**: the affected scope now reads the widget names of every changed page/snippet and of each calling page (`DESCRIBE PAGE`, one extra warm-catalog `mxcli` call, 0.07 s) and selects each journey whose JSON holds one; a navigation or layout change selects every journey; when callers or widgets cannot be measured (`widgets NOT MEASURED` row) it selects every journey (unsure means include). Field: 067 (Request_Edit page) 0 -> J-REQ-01, 090 (navigation) 0 -> all 5 journeys — speed experiment E2
+- fix(harvest-learnings): **a bug-log heading the toolkit log already carries is no longer drafted again when it has `~`, `*` or a backtick in its first 30 characters after the `BUG-nn:` tag.** Section 1 strips those characters from a project heading before it matches the heading's first 30 characters after the tag against `bug-logs/mxcli-bugs.md`, but searched the log as written, so the stripped text was never a substring of a heading like ``BUG-01: `alter entity drop attribute` causes …``. The log is now stripped the same way before the search, and the search takes `-e`, because stripping can leave a heading that starts with `--` (BUG-26, BUG-54), which `grep` read as an option. Reproduced with the script at 95e1b30 on a project whose bug log holds copies of all 223 headings in the toolkit log: it drafts 103 of them, and 0 after the fix. Pinned by `tests/wave2/test-harvest-learnings.sh` T6 (4 headings taken from the log itself; T6 fails on 95e1b30) — saingge
 - fix(canvas): **`mf-canvas-facts.py` reads a note line's sides by the model's Mendix version, so on Mendix 11 it no longer reports NOTE-SIDE for lines that face their object.** A note numbers its sides top 0, bottom 1, left 2, right 3 on Mendix 10, but like every other object (top 0, right 1, bottom 2, left 3) on Mendix 11; the script knew only the Mendix 10 table. It now reads `_ProductVersion` from the `.mpr`'s `_MetaData` table (major 11 or higher: the activity numbering; unreadable: Mendix 11) and prints `# note sides: …` as its first line. One app's 305 hand-drawn note lines, matched across its 10.24 and 11.15 versions: the upgrade renumbered the note's side on every line (0 to 0, 1 to 2, 2 to 3, 3 to 1) and left the object's side alone, and the script's fact counts for the whole app are now identical on both. mxcli v0.28.0 (mendixlabs/mxcli#1620) writes the activity numbering whatever the project's version. Fixture section V (renumbered copy of the golden model). — saingge
 - process(design-artifacts, conversion-runbook): **Stage 3 runs in parallel by default, from one brief and a mechanical task list**: the lead fixes modules, canonical names and design tokens, writes `docs/progress/stage3-brief.md`, and takes the tasks from the stage-3 lines of `bin/lib/artifact-check.sh` (so no artifact is forgotten); one subagent per 1–2 screens or artifact group, a shared opening line so the start is cached, mechanical tasks on the cheap tier, 5-line replies, the lead the only writer of `PROJECT.md`. The runbook's Stage 3 fan-out row points there, and its Stage 2 row names `facts-lock.sh resolve`. Measured on one project (single runs, Sonnet, 250k auto-compact): serial 13 min / $3.60, parallel 10 min / $4.44, parallel with a brief 6 min / $4.32; 0 name mismatches in the parallel runs (serial dropped 3 entities); both parallel runs forgot the process map and swimlanes, which the task list now covers. Baseline words 64,004 before and after — Stage 3 experiment 2026-10-28
 - fix(facts-lock, gate-check): **parallel BRDs keep one spelling per identifier**: `bin/facts-lock.sh <project> resolve` freezes each casing conflict (`WIPStatus`/`WipStatus`) whose majority spelling is used by strictly more BRD sources than any other and records the count in the lock (`resolvedBy`); a tie stays contested, and `build` still never picks on its own. Stage 2 of `gate-check.sh` now reads the lock once there are 2+ BRDs: a BRD that contradicts a frozen identifier FAILs and is quoted; no lock yet is a note, never a FAIL. Measured cause: 20 parallel BRDs on PROJECT-C carried 10 casing conflicts into Stage 3 as open questions, and facts-lock.sh, documented only as advice, reached 1 of 6 projects. Fixtures: `test-facts-lock.sh` resolve cases (4), `test-bug03-gates.sh` T19 (2 of 4 fail on the previous gate-check) — speed/parallel experiments 2026-10-28

diff --git a/bin/harvest-learnings.sh b/bin/harvest-learnings.sh
@@ -70,6 +70,11 @@ BUGS_OUT="$INBOX/$STAMP-$PROJECT_NAME-bugs.md"
 FOUND_BUGS=0
 if [ -d "$PROJECT_DIR/bug-logs" ]; then
   TMP="$(mktemp)"
+  # A heading is stripped of ~ * ` before it is matched (below), so the log it is matched against is
+  # stripped the same way. Without that, a heading the toolkit log carries with a backtick in its
+  # first 30 characters never matched, and was drafted again.
+  BUGLOG_PLAIN="$(mktemp)"
+  tr -d '~*`' < "$TOOLKIT_BUGLOG" > "$BUGLOG_PLAIN" 2>/dev/null
   find "$PROJECT_DIR/bug-logs" -name '*.md' -type f | sort | while IFS= read -r f; do
     # Split the file on '## ' headings; test each heading's distinctive text against the
     # toolkit log. awk emits candidate sections whole, delimited for the outer loop.
@@ -86,7 +91,8 @@ if [ -d "$PROJECT_DIR/bug-logs" ]; then
       # measured: personal-toolkit and the toolkit disagree from 82 up), then take 30 chars.
       slice="$(printf '%s' "$heading" | sed -E 's/^BUG-[0-9]+[: ]*//' | cut -c1-30)"
       [ -n "$slice" ] || continue
-      if ! grep -qiF "$slice" "$TOOLKIT_BUGLOG" 2>/dev/null; then
+      # -e: stripping can leave a slice that starts with "-", which grep would read as an option.
+      if ! grep -qiF -e "$slice" "$BUGLOG_PLAIN" 2>/dev/null; then
         { printf '## [candidate — from %s] %s\n' "${f#$PROJECT_DIR/}" "$heading"
           printf '%s' "$section" | tail -n +2
           printf '\n'; } >> "$TMP"
@@ -99,7 +105,7 @@ if [ -d "$PROJECT_DIR/bug-logs" ]; then
     FOUND_BUGS=1; WROTE=$((WROTE + 1))
     echo "  wrote ${BUGS_OUT#$TOOLKIT_ROOT/}"
   fi
-  rm -f "$TMP"
+  rm -f "$TMP" "$BUGLOG_PLAIN"
 fi
 [ "$FOUND_BUGS" -eq 1 ] || echo "  bug-logs: nothing the toolkit log doesn't already carry (or no bug-logs/)"
 

diff --git a/tests/wave2/test-harvest-learnings.sh b/tests/wave2/test-harvest-learnings.sh
@@ -25,6 +25,8 @@
 #   T5  section 1 drafts every '## ' section of every bug-log file, not only the first of each,
 #       and still skips a later section whose heading the toolkit log carries. Fixture:
 #       fixtures/harvest-learnings/ (see its CAPTURE.md).
+#   T6  a heading the toolkit log carries is not drafted again when it has ~ * or ` in its first
+#       30 characters, or starts with a dash once those are stripped.
 #
 # Usage: bash tests/wave2/test-harvest-learnings.sh /path/to/harvest-learnings.sh
 
@@ -196,8 +198,7 @@ else
   EXPECTED5="$(cat "$FIX5"/bug-logs/*.md | grep '^## ' | sed 's/^## *//; s/[~*`]//g')"
   N5="$(printf '%s\n' "$EXPECTED5" | grep -c .)"
   # A later section whose heading the toolkit log already carries must still be filtered out.
-  # Taken with no ~ * or ` in it: the harvester strips those from a heading before matching it
-  # against the log, so a heading that has them never matches and is drafted every time.
+  # A plain heading is enough here; T6 covers the ones with ~ * or `.
   KNOWN5="$(grep -m1 -E '^## BUG-[0-9]+: [^~*`]+$' "$CLONE/bug-logs/mxcli-bugs.md")"
   if [ -n "$KNOWN5" ]; then
     printf '\n%s\n\nA section the toolkit log already carries; never drafted.\n' "$KNOWN5" \
@@ -235,6 +236,41 @@ else
   fi
 fi
 
+# ---------------------------------------------------------------------------------------------
+# T6: a heading the toolkit log carries is not drafted again when it has ~ * or ` (or starts with
+# "-" once stripped). The harvester strips ~ * ` from a project heading before matching it, but searched the raw
+# log, which keeps them, so the stripped slice never matched: 103 of the log's own 223 headings were
+# drafted again from a project that held nothing but copies of them. A heading that starts with a
+# backtick and then "--" starts with "--" once stripped; grep read the slice as an option, which
+# also counted as "not found". The headings come from the
+# clone's own log, so the check follows the log instead of a copy of it.
+# ---------------------------------------------------------------------------------------------
+echo "== T6: headings with ~ * \` or a dash left first are recognised as known =="
+
+if [ ! -f "$CLONE/bin/harvest-learnings.sh" ] || [ ! -f "$CLONE/bug-logs/mxcli-bugs.md" ]; then
+  bad "control: no isolated clone with bug-logs/mxcli-bugs.md (see T1) — fixture cannot run"
+else
+  P6="$WORK/knownproj"; mkdir -p "$P6/bug-logs"
+  { grep -m3 -E '^## BUG-[0-9]+: .{0,26}[~*`]' "$CLONE/bug-logs/mxcli-bugs.md"
+    grep -m1 -E '^## BUG-[0-9]+: [~*`]-' "$CLONE/bug-logs/mxcli-bugs.md"; } \
+    | awk '{print; print ""; print "Copy of a heading the toolkit log carries."; print ""}' > "$P6/bug-logs/known.md"
+  N6="$(grep -c '^## ' "$P6/bug-logs/known.md")"
+  if [ "$N6" -lt 4 ]; then
+    bad "control: the clone's log gave $N6 of the 4 headings this check needs (3 with ~ * \` in their first 30 characters, 1 that starts with a dash once stripped)"
+  else
+    printf '\n## BUG-900: a heading no toolkit log has\n\nGenuinely new.\n' >> "$P6/bug-logs/known.md"
+    OUT6="$(bash "$CLONE/bin/harvest-learnings.sh" "$P6" --to "$WORK/inbox6" 2>&1)"; RC6=$?
+    DRAFT6="$WORK/inbox6/$STAMP-knownproj-bugs.md"
+    [ "$RC6" -eq 0 ] && ok "exit 0" || bad "exit $RC6, expected 0" "$OUT6"
+    GOT6="$(grep '^## \[candidate' "$DRAFT6" 2>/dev/null)"
+    if [ "$(printf '%s\n' "$GOT6" | grep -c .)" -eq 1 ] && printf '%s' "$GOT6" | grep -qF 'BUG-900: a heading no toolkit log has'; then
+      ok "$N6 known headings (with ~ * \` in them, or a dash left first) not drafted; the new one is"
+    else
+      bad "expected only BUG-900 drafted; got:" "$GOT6"
+    fi
+  fi
+fi
+
 echo ""
 echo "PASS=$PASS FAIL=$FAIL   ($WORK)"
 # Every draft this fixture produced lives under $WORK (the clone's contrib/inbox, or the no-git