[MendixMau]

Direct lane

Classification: DISCUSS. The maintainer approved this as option C of the lazy-loading plan. PR 1 is #570.

Split. The whole of PR 2 changes about 630 hand-written lines plus 2,634 generated lines, so it is split per the brief:

• 2a (this PR): the lookup script, the kind table, the generated files, the fixture and the CI check. No routing change.
• 2b (branch claude/loving-gates-hauc1n-lazy-pr2b, stacked on this one, pushed, not opened yet): the routing demotion, the gate-check fix, the mdl stub line and its fixture cases.

The word counts and the A/B below cover both halves, because the A/B arm L is the 2b preload.

What changed and why. Today the mdl helper preloads every kind-specific MDL skill at Stage 5: 29,388 words. A page script needs none of the microflow rules, and a microflow script needs none of the page rules. This PR adds the per-statement-kind lookup.

• bin/mdl-rules.sh <kind...> | --for <script.mdl> | --render | --check. It prints the always-needed core, then the sections that bin/lib/mdl-rule-kinds.tsv maps to each kind (page, microflow, access, workflow, entity, any). It ends with Rules for kinds K: STOP rows n/n listed. The denominator is counted by a second, independent pass over the STOP table. If the two counts disagree, the run fails.
• Sections are located by file and heading text, never by line number (the runbook-span.sh precedent). The table covers:
  • the STOP rows, filtered by PR 1's Kind column
  • the kind-specific gotchas (per bullet)
  • learned-microflow-patterns.md
  • microflow-preflight.md
  • ui-preflight-pages.md
  • design-spacing.md §1–3
  • the detection-gaps register rows, by an ERE over the construct cell
  • security-is-not-a-later-script.md
  • retesting-learned-rules.md
• --for reuses the existing classifier by calling project-bin/exec.sh --subjects. That is the same lexer exec.sh, affected-tests.sh and the change class use. There is no second copy, and exec.sh is not refactored.
• It never prints empty. A missing heading, a bullet lead that matches nothing, an ERE that matches no row, or a STOP count mismatch each give exit 1, a message on stderr and nothing on stdout.
• skills/generated/rules-<kind>.md hold the same bytes as bin/mdl-rules.sh <kind>, for Copilot, Cursor and Windsurf, which read files. They have a routing-exempt.tsv line: they are rendered output, not routed one by one.
• --check runs in CI (new step next to render-routing.sh --check). It checks the denominators: every STOP row has a known Kind, every gotcha bullet and every register row reaches at least one kind, and the generated files match a fresh render. A new gotcha or register row that no kind selects fails CI instead of silently not reaching the mdl helper.
• tests/behaviour/build-packs.sh --rules "<kinds>" expands the script in a pack to its output for those kinds. That is how the dispatcher will deliver it in PR 3.
• No skill text changes. Bash 3.2 and POSIX awk. It makes no mxcli or mxbuild calls.

Row	Answer
Process	2a: nothing new blocks a project; CI gains bin/mdl-rules.sh --check. 2b: the mdl stub gains workflow step 4a: run bin/mdl-rules.sh (or read rules-<kind>.md) before drafting, and report "STOP rows for kinds K: n/n checked" in PREFLIGHT.
Dependencies	It reads project-bin/exec.sh --subjects, and its output is unchanged. Producer: a toolkit checkout; the script needs no project. 2b touches skill-routing.tsv/.sh, gate-check.sh and agents/mdl-agent.md.
Context cost	2a: none (the generated files are exempt and not routed); baseline stays at 72,891. With 2b (render-routing.sh --check, LC_ALL=C): baseline 72,891 → 63,877 (−9,014, the three mdl-only files go on-demand). mdl Stage 5 29,388 → 14,600 words / 14 → 7 files. The 14,600 counts the 3,347-word core of mdl-rules.sh, like the runbook slice. Per task the helper adds its kinds on top: page 10,223, microflow 10,186, access 4,730, entity 5,183, workflow 3,732 (each includes the core). So a page task reads about 21.5k words instead of 29.4k. This is above the plan's 12–13k estimate because the core keeps Step 0, the 3 any STOP rows, the gotchas, the final self-check, the detection-gaps ladder and operating rules, and retesting-learned-rules.md whole. Other roles do not change.
Behaviour	A/B below. The decision rule as written is not met: there are three single-rep losses, reported, not averaged away.
Runtime	mdl-rules.sh page takes about 0.1 s; --check takes about 1 s.

Gate / sync fix (in 2b)

gate-check.sh build-ready item 2 grepped CLAUDE.local.md for ui-preflight-pages. Once that row is on-demand, a re-synced baseline block no longer names it, so item 2 would flip to FAIL. 2b accepts ui-preflight-pages or bin/mdl-rules.sh, which is the new baseline row.

The sync-project.sh "§2c" check the brief mentions no longer exists. §2d now renders the routing block instead of grepping for it (see the T9 note in test-bug12-sync.sh), so there is nothing to change in sync. The only check is the gate's.

Fixture test-mdl-rules.sh T7 (2b) covers:

• a freshly synced project, single-tree and two-tree (.mpr under app/): item 2 passes, and the rendered block holds bin/mdl-rules.sh and no ui-preflight-pages
• an old, unsynced table: still passes
• a file with neither name: still fails

Fixtures (scoped runs only; full suites not run)

• test-mdl-rules.sh bin/mdl-rules.sh: 26/26 on 2a, 30/30 on 2b. It covers:
  • each kind (own sections, none of another kind's, n/n, byte-equal to the generated file, core present)
  • --for on the real mixed page+entity repro pipelines/mendix-mpr/samples/.../D2-page-repeater-content.mdl
  • --for on real workflow, access, microflow and scheduled-event scripts
  • an unknown kind (exit 2, empty stdout)
  • a renamed heading and a dead bullet lead (exit 1, empty stdout, named on stderr)
  • --check drift, an unmapped new gotcha, an unmapped register row and an unknown STOP Kind (all on a temp copy)
• Golden input is the real skills and the repo's real .mdl files; nothing is hand-written.
• Run on 2b: test-mdl-scan 52/0, test-baseline-pack 22/0, test-render-routing 5/0, test-routing-tier 26/0, test-runbook-slice 31/0 (T5 now counts mdl-rules.sh as its core output), test-bug03-gates 47/0, test-source-sufficiency-gate 7/0, test-source-ledger 55/0, test-bug12-sync 62/0, test-authoring-audit 5/0.
• render-routing.sh --check is clean. The authoring-audit.py routing advisory is unchanged (top-1 49/55, top-3 55/55). check-portability and check-scripts are clean.

Behaviour A/B (tests/behaviour/RUNBOOK.md)

• Arms. M = origin/master @76e1610, pack 33,296 words. L = 2b head @785bef2: the mdl Stage-5 preload, with bin/mdl-rules.sh expanded to its output for the task's kinds via the new build-packs.sh --rules mode. L packs are 19.9k–26.7k words.
• Kinds per task:
  • page task: page, access
  • mf task: microflow, access
  • trap 12: page
  • traps 26, 1b: page, access
  • traps 33, 29: microflow, access
  • trap 16: access
  • trap 18: workflow, microflow, access
• Setup. Fresh claude -p agents with only Read and Write, model claude-opus-5-5, run from an empty directory with no CLAUDE.md, arms interleaved.
• Reps. 5 page + 5 mf per arm (3 planned, plus 2 more to separate noise from loss), 1 per trap per arm, and 2 for trap 18.
• Traps. STOP 4 is planted in the page task and STOP 21 in the mf task (score.sh planted). PR 1's traps cover 12, 26, 1b, 33 and 29. New traps: 16 (access XPath with a path inside [%CurrentUser%]) and 18 (callee microflow and workflow in one script).

Metric	M	L
read_all	30/30	30/30
page: sections, report, scan, planted (STOP 4), crumbs, txtcrumb, h1, empty, folder	equal in every rep	equal
page: quoted (median)	100 (62–100)	100 (all 100)
page: uixref	5/5	4/5 — LOSS (L-page-r3 posted Step 5 as NOT RUN lines but no "UI cross-reference" block)
mf: sections, report, scan, planted (STOP 21), cinloop, rinloop, pos, vfb, folder	equal	equal
mf: quoted (median)	90	100
mf: mfpre	5/5	4/5 — LOSS (L-mf-r3 posted both checklists with tiers but wrote microflow-preflight — where the scorer wants microflow-preflight:)
traps 12, 26, 1b, 33, 29, 16: planted / flagged	1/1 each	1/1 each
trap 18: planted (mechanical)	2/2	0/2 — LOSS by the scorer
tokens (mean page / mf)	323k / 311k	314k / 272k

About trap 18. Both L reps did split into two scripts (scripts: 2, "FILE 1 of 2 / FILE 2 of 2", STOP #18 cited). They wrote both files inside the single ```mdl fence the task format demands, and the ad-hoc scorer counts per fence. Both M reps left the workflow undrafted. Read by hand, both arms obey STOP 18. Read mechanically, L loses.

Verdict. Under the decision rule ("no worse on every metric, count at the good value"), L is not shown no-worse. There are three losses, each a single rep, and two of them are output-format misses where the content was present. This goes to the maintainer to decide.

Outputs and scores are in the builder's session scratch. They can be reproduced with build-packs.sh and the PR 1 trap tasks plus the two new ones.

Field evidence. None. A field run is owed: this was not run on a real project. It is planned with PR 3, the dispatcher that hands mdl-rules.sh output to the helper.

• No client data anywhere in the diff
• Size cap: 2a is about 430 hand-written lines plus 2,634 generated lines (skills/generated/, rendered output) — split 2a/2b per the brief
• Test tier: scoped wave2 fixtures (listed), plus the behaviour A/B
• Golden input captured (real skills, real .mdl), both layouts (T7 in 2b); Bash 3.2-safe constructs and Git Bash considered (no .exe probing needed, only bash and awk), but not run on macOS or Windows
• CHANGELOG.md line in this PR (and one in 2b)

Open questions for the maintainer

• This repo's CLAUDE.md still names learned-microflow-patterns.md in the "always-on set". After 2b it is on-demand. That one-line edit was left to the maintainer.
• bin/init-project.sh item 5 and project-bin/context-pack.sh still point at the skill files directly. They are lead/dispatcher surfaces and left for PR 3.

Review: pending independent review

---

Generated by Claude Code