[MendixMau]

Direct lane

What changed and why (one paragraph):
PR 1 of the "Option C" lazy-loading plan. It moves text and adds an index. Only skills/learned-mdl-preflight.md changes, plus its destination and the CHANGELOG. What moved to process/baseline-incident-history.md § PF1 (new section, in that file's existing format, verbatim):

• the "Re-probed on…" line and the old "Version context" line;
• the whole Version notes (mxcli ≥ v0.20.0) block;
• the STOP table as it stood before the trim, all 34 rows verbatim. That includes the 15 retired rows (1c 2 3 5 6 6b 7 8 9 10 11 13 15 17 27) and every live row's full root cause.

In the skill:

• The retired rows collapse to one line that lists their numbers and points to § PF1.
• Each live row keeps its If and Then cells unchanged, plus one short failure clause in a "Fails as" column.
• A new Kind column sits at the end of each row: page | microflow | access | workflow | any. No live row came out as entity; row 23 covers documents and entities, so it is any.
• Pointers to § PF1 sit where the text moved: the Version context line, a line above the table, and the retired-rows line.

Protected text that stays:

• every live row's If/Then text, verbatim;
• row 1's safe CLI alternative, its ContentParams ban and per-value colours, which were in the root-cause cell (they are rewrites);
• 1b's "read the page back with describe page";
• 22's "the exit code is the only signal: never ignore it";
• 23's BUG-146 "check each - Attribute against DESCRIBE ENTITY";
• Step 0, all MDL gotchas, the final self-check and every denominator;
• the workflow DECISION STOP (BUG-76, "never force with --no-check") and the #879 "read the ledger entry before a first marketplace install" pointer. Both used to sit in the version notes and are kept as one line.

Parsers kept working:

• test-mdl-scan.sh:56: the scan key dynamictext (Attribute: "State") is only in this file, not the ledger. It stays in row 1.
• :90: the | <n> 🔎 markers still match exactly 1/12/16/26. The Kind column sits at the row end, and the retired line starts | 1c 2 … with no 🔎.
• No other parser in bin/ or tests/ reads the table's content. sync-project.sh, wiring-item3.sh, gate-check.sh and the bug06, bug12, sync-dup and context-audit fixtures only use the file path.

Impact

Row	Answer
Process	none: no gate or step added or removed
Dependencies	skills/learned-mdl-preflight.md, process/baseline-incident-history.md, CHANGELOG.md. No shared script touched.
Context cost	render-routing.sh --check baseline 73,946 → 72,891 words (−1,055), budget 80,000; routing surfaces in sync. Stage 5 mdl role 30,443 → 29,388; the other stages and roles are unchanged. Below the ~−1.3k estimate because the If/Then cells, which hold most of the remaining words (rows 16, 21, 22, 23, 24), are protected.
Behaviour	Reduced A/B, M = origin/master 888b4b2, T = 4f5734e. 3 page + 3 mf per arm, plus 1 rep per arm of 5 new planted traps (STOP 12, 26, 1b, 33, 29): 22 reps. Every rep was a fresh claude -p (claude-opus-5-5, Read/Write only) run from an empty directory with no CLAUDE.md. T no worse than M on every metric; no losses. Details below.
Runtime	n/a

Moved-text check: bin/check-moved-text.sh process/baseline-incident-history.md skills/learned-mdl-preflight.md found 43 of 43 removed non-blank lines verbatim in the destination (exit 0).

Fixtures (scoped): tests/wave2/test-mdl-scan.sh: 52 passed, 0 failed. bin/render-routing.sh --check: in sync. The full suites were not run.

A/B detail:

• All 22 reps read their pack to the end (read_all=1), so there were no partial reads.
• page, M and T alike: sections 4/4, report 1, planted 1/1/1, scan 0, crumbs 2/3, h1 2/2, empty 1, uixref 1, folder 1/3.
• page quoted: M 100/92/95, T 100/100/100.
• mf, M and T alike: planted 1/1/1, mfpre 1, cinloop/rinloop/pos 0, vfb 1, scan 0.
• mf quoted: M 89/84/88, T 90/90/87.
• check = NA in both arms (no mxcli binary).
• Traps (planted / flagged): 12, 26, 1b, 33 and 29 each scored 1/1 in both M and T.

The trap tasks and their scorer were ad hoc. They were not added to tests/behaviour/, because score.sh scores a planted request only for page and mf, and adding tasks would widen this PR. A run used 7 turns in both arms. Tokens per run: page M 384–482k, T 340–396k; mf M 258–480k, T 356–462k.

Field evidence: n/a. Text only; no instrument changed.

• No client data anywhere in the diff
• Size cap: 3 files, 78+/44−
• Test tier: scoped fixture + render-routing check + behaviour A/B (reduced)
• CHANGELOG.md line appended in this PR

Review: pending independent review

---

Generated by Claude Code