skills: decompose-app, rehost-app + mendix-service-seam (split from #293, 4/6)

[MendixMau]

Split from #293 (part 4 of 6).

Stacked on #342 (split-293/rehost). The skills cite pipelines/mendix-mpr/, so #342 must merge first. Until then this diff also shows #342's first commit. This branch does not yet contain #342's review-fix commit 300cd35; it will be merged in with master once #342 lands. This PR's own commits are 3ac141e and 3c472fe.

Depends on the split-293/cut-score and split-293/cut-ledger PRs. decompose-app.md names bin/cut-score.sh (Stage 1) and bin/cut-ledger.sh (Stage 4, Stage 6). Neither file is on master yet.

Direct lane

What changed and why (one paragraph):
This adds three on-demand skills from #293:

• decompose-app.md: split a Mendix monolith into several apps. It covers:
  • picking the cut on business fit and coupling
  • a gated target architecture
  • cut order by dependency direction
  • moving each slice by MPK
  • ledgering every crossing edge
  • reporting exactly one level per slice
• rehost-app.md: regenerate a whole .mpr into a fresh app with pipelines/mendix-mpr/ and read the fidelity ledger. Done means build-green with named hand-offs plus journey parity.
• mendix-service-seam.md: OData and REST between two Mendix apps, as patterns P1–P13, the external-entity limits, recipes A and B, and completion criteria.

Each skill has its row in bin/lib/skill-routing.tsv, re-rendered with LC_ALL=C bin/render-routing.sh. That updated README, ROUTING, three agent stubs and the gate-check.sh stage map. architecture-blueprint.md and learned-constants-and-secrets.md gain the one-line pointers from #293, and the pipeline README now links to the skills.

Changes from #293:

• Inbox citations. All contrib/inbox/ file citations are rewritten: 21 file mentions in Decompose and Rehost pipelines: field runs on FSM and GenAI Showcase, three new skills #293's text.
  • The decompose-app traps now name the increment that measured each trap, and the Evidence header defines the increments.
  • The rehost-app Step 6 pointer now goes to the pipeline README.
  • The remaining mentions of contrib/inbox/ are the contribution lane as a destination for new findings. None cites a file.
• Other dead citations rewritten the same way: samples/… captures and FINDINGS files, the process plan, a Decompose and Rehost pipelines: field runs on FSM and GenAI Showcase, three new skills #293 commit sha, web-slice-candidates.md and cut-score-findings.md. Each now states the fact with its run.
• App names replaced by app type.
• Routing rows rewritten as trigger conditions (authoring rule 1).

Authoring rules 1–8:

1. Trigger rows: done, as above.
2. Baseline-tested against fresh agents: not done.
3. Prohibitions vs recipes: the skills give recipes for shape problems and bans for discipline problems.
4. Denominators: every step starts with a Bound: line with a checkable count.
5. No cached environments: the skills point to the actual script headers.
6. One skill per situation: the three are cross-linked with "Not this skill" lines.
7. Failure shown next to each rule: the patterns pair the rule with the verbatim failure.
8. Retrieval: not tested beyond render-routing --check's coverage audit.

Field evidence (from #293's runs, all on public apps):

• Field-service template (Mendix 9.24.2), increments 2–8 on 2026-10-12/13:

  • read probe and cut scoring
  • MPK moves
  • a native notes slice
  • an asset/parts slice made seam-live, then strangled: in-edge hand-offs 499 → 157, journey 16/16 twice
  • a dispatch/workforce slice run from these skills: stub rows 271 → 3

• Increment 1: a synthetic 4-module app (Mendix 10.24).

• GenAI showcase app, rehost: 36/36 modules landed, not build-green.

• Banking sample, rehost: build-green.

• Not verified: macOS, and the two-tree layout (.mpr under app/).

• No client data anywhere in the diff (leak guard clean; whole-word denylist grep clean)

• Size cap: three skills, about 590 lines of prose. Over the 400-line cap. Each skill could ship as its own PR.

• Test tier reached: T0 plus the scoped routing checks:

  • render-routing --check: in sync
  • test-render-routing.sh 5/0, test-routing-tier.sh 12/0, test-baseline-pack.sh 16/0, each run with bin/render-routing.sh
  • check-scripts, check-portability, check-no-client-data, check-no-private-citations, check-pr-discipline all clean

• New/changed instrument: n/a (prose only)

• For a new skill: routing rows added and re-rendered

• CHANGELOG.md line appended, credited

• No bug entries

Review (MERGE-AFTER-FIX), addressed in 3c472fe

1. MEDIUM: wrong increment number. The cut-score blind spots in decompose-app.md Stage 1 step 4 were measured in increment 2, the first cut scoring on the field-service template. They were wrongly cited as increment 5. The Evidence header now also covers the first two increments:
   • increment 1: bin/cut-score.sh built on a synthetic 4-module app (Mendix 10.24)
   • increment 2: read probe and first cut scoring on the template
2. LOW: level amendment dropped. Stage 6 now says the three levels were CONFIRMED with the user 2026-10-12. It also records that level 3 was amended 2026-10-13 from field evidence: "one journey per crossing edge" became one journey per crossing kind, because per edge was undefined for 536 in-edge rows (increment 7). The level 3 row carries the amended criterion.

Checks re-run after the fix: render-routing --check in sync, the three routing fixtures 5/0, 12/0 and 16/0, and all guards clean.

Baseline word count: 75,038 before, 75,038 after (budget 80,000).

Impact

Row	Effect	Evidence
Process	None new at gate-check level: three on-demand skills; nothing new blocks a gate. The skills define per-slice levels and steps for the decompose and rehost paths.	diff
Dependencies	Shared files: bin/lib/skill-routing.tsv and its re-render (README.md, ROUTING.md, agent stubs architect/ba/mdl, bin/gate-check.sh stage map), CLAUDE.md, skills/architecture-blueprint.md, skills/learned-constants-and-secrets.md. Merges clean onto master 868b3cb (git merge-tree). Conflicts with #342 (add/add, pipelines/mendix-mpr/*; stacked on an older #342 head) and with #273 (bin/gate-check.sh stage map). Order: #341, #345, #342, then this PR rebased and re-rendered. Cites bin/cut-score.sh (#341), bin/cut-ledger.sh (#345) and pipelines/mendix-mpr/ (#342), none on master yet.	git merge-tree vs 868b3cb and pairwise
Context cost	Unchanged: baseline 73,522 → 73,522 words (budget 80,000); Stage 5 mdl 30,185 → 30,185, lead 25,820 → 25,820	bin/render-routing.sh --check on master 868b3cb vs this head (3c472fe) merged onto 868b3cb via git merge-tree in a temporary worktree
Behaviour	A/B REQUIRED, NOT RUN. Three agent stubs gain rendered routing rows. Situational text also changes (three new on-demand skills, architecture-blueprint.md, learned-constants-and-secrets.md): A/B recommended.	diff of agents/*.md; routing rows load class ondemand
Runtime	None (stage-map data only in gate-check).	diff

🤖 Generated with Claude Code

https://claude.ai/code/session_01VJgWP5vEoAsNsJYqCDMGNw