decompose: cut-score + app-facts fix (split from #293, 1/6)

[MendixMau]

Split from #293 (1/6). #293's branch is untouched; files were copied from its head with git checkout origin/claude/youthful-hypatia-qapq73 -- <paths> and adapted to current master.

Direct lane

What changed and why (one paragraph):
bin/cut-score.sh scores a proposed split of an existing app into several apps. You pass --cut "A,B|C|D", or it uses the default of one app per level of app-layer-map.json. It lists every module edge in analysis/app-facts/dependencies.json that crosses an app boundary and names the integration lane each ref kind would need: OData external entity, published action or REST, copy, deep-link, or none. It then gives each partition a clean / costly / blocked verdict and prints the denominator. Aggregates over a foreign entity are read from .mxcli/catalog.db. When that can't be done, rows read NOT DETECTED and name the actual reason: no catalog, empty table, missing columns, or unreadable. The fixture is tests/wave2/test-cut-score.sh plus the captured tests/wave2/fixtures/cut-score/. The project-bin/app-facts.sh change from #293 runs the one-column guard queries through Python's sqlite3 when the sqlite3 CLI is missing, stripping the \r that native Windows Python prints.

Adapted to master (not in #293): #259 changed the catalog guard to stop testing the build_mode label, because mxcli v0.24+ labels a complete catalog source. #293's cut-score still tested build_mode != 'full'. Now the label is recorded and not tested, and an empty activities_data reads NOT DETECTED. The header no longer cites the decompose plan doc, which lands in a later split; it now points at skills/layering-review.md.

Field evidence: this is #293's run, now stated in the script header.

• App: a public Mendix field-service template app (Marketplace, 9.24.2, MPR v1), mxcli v0.26.0, 42 own modules, 312 own edges, 4,508 refs. app-facts took 25 s; each cut-score run took ≤ 1 s.

• Cuts scored: seven, all blocked. One app per level crossed 312 edges / 4,508 refs, which means little on its 36-module tangle. The best web | mobile landscape found was 132 edges / 1,442 refs with 2 bidirectional pairs.

• Spot-check: 6 blocker rows checked with describe. All 6 were real references, but 2 had the wrong lane.

• Captured fixture: a planted four-module Mendix 10.24 app (fixtures/cut-score/CAPTURE.txt).

• Known limits from that run, not fixed here:

  • call mixes client-side nanoflow/JS calls in with server-side ones. That is about 30% of the main seam's blocker mass.
  • The catalog kinds sync, widget, schedule, event and publish never reach dependencies.json.
  • Nanoflows that aggregate the device store produce false aggregate hits.

• No client data anywhere in the diff (leak guard and whole-word name grep are clean)

• Size cap: exceeded. This PR needs the maintainer's size waiver. There are 612 non-test lines: bin/cut-score.sh (~600) plus app-facts (+12). I looked at splitting off the HTML renderer (~180 lines). It is not a clean cut: HTML is the default output, -o and fixture section 7 depend on it, and what remains would still be about 430 lines, over the cap.

• Test tier reached: T1, fixtures on captured input plus synthetic layout trees and derived probes:

  • test-cut-score.sh 70/70. That is Decompose and Rehost pipelines: field runs on FSM and GenAI Showcase, three new skills #293's 56 plus 14 new checks: the source label, an empty table, row reasons, the own-module derivation, and single-tree and two-tree catalog discovery. Against Decompose and Rehost pipelines: field runs on FSM and GenAI Showcase, three new skills #293's unadapted copy, 8 of the new checks fail.
  • test-app-report.sh 109/109.
  • test-catalog-edge-guard.sh 22/22, including a new case that forces the Python path with CRLF output. Without the tr it fails 2.

• For a new/changed instrument: golden input is captured (CAPTURE.txt). Two-tree layout (.mpr under app/): probed only with synthetic trees, and cut-score finds the catalog in both. No real two-tree project has been run. macOS and Git Bash/Windows: unverified. Everything ran on Linux. The CRLF case simulates native Windows Python's output, but nothing has been run on Windows.

• CHANGELOG.md line appended (crediting Decompose and Rehost pipelines: field runs on FSM and GenAI Showcase, three new skills #293's field runs)

Checks: render-routing.sh --check passes; the baseline is 75,038 words before and after. check-scripts, check-portability, the leak guard, check-no-private-citations and check-pr-discipline are all clean.

Review

The independent review returned MERGE-AFTER-FIX. Commit c55182f addresses each item:

1. MEDIUM: the app-facts Python fallback did not strip \r, so on native Windows Python every [ "$(q …)" = "1" ] failed with "catalog schema is missing".
   • Fixed: | tr -d '\r', as _common.sh does.
   • Test seam: APPFACTS_NO_SQLITE3_CLI=1 forces the Python path.
   • New fixture case: in test-catalog-edge-guard.sh, a Python wrapper that ends every line in CRLF, run on both layouts. It passes 22/22 with the fix and 20/22 without it.
2. LOW: rows said "NOT DETECTED (no catalog)" even when a catalog existed. Fixed: they now carry the actual reason, with new cases for an empty table and an unreadable file.
3. LOW: in {from} | {to} - {None} the precedence was wrong. Fixed: parenthesised, with a new case of no scope plus an edge missing from. Decompose and Rehost pipelines: field runs on FSM and GenAI Showcase, three new skills #293's copy crashes on it.
4. Size: waiver requested; the reasoning is in the size-cap item above. Field proof: the template-app run is now stated inline in the header, as results only.

Impact

Row	Effect	Evidence
Process	None new. bin/cut-score.sh is an on-demand instrument, not wired into any gate. app-facts.sh no longer stops when the sqlite3 CLI is missing (falls back to Python's sqlite3).	diff
Dependencies	Shared file: project-bin/app-facts.sh (installed script). Merges clean onto master 868b3cb (git merge-tree). No conflicts with the other PRs in this batch; #346 cites bin/cut-score.sh, so this lands before #346. Platform: Linux only; native-Windows CRLF path simulated, not run on Windows or macOS. Reads analysis/app-facts/dependencies.json (producer: project-bin/app-facts.sh), analysis/app-layer-map.json (producer: bin/app-layer-map.sh; without it --cut is required), optional .mxcli/catalog.db (left by app-facts.sh). Size waiver requested.	git merge-tree vs 868b3cb and pairwise; cut-score.sh header
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 (c55182f) merged onto 868b3cb via git merge-tree in a temporary worktree
Behaviour	n/a (scripts only).	diff touches no skill or agent stub
Runtime	No change to exec, gates or checks. cut-score ≤ 1 s per run on the field app (42 modules); app-facts unchanged on the CLI path.	PR field-evidence section

🤖 Generated with Claude Code

https://claude.ai/code/session_01VJgWP5vEoAsNsJYqCDMGNw