deploy

name	deploy
description	Historical deploy notes. Current Radia deploy/release uses release-qud plus cubit-plugin-install; do not follow old .ccl/Cubit-2025.3 recipes except for archaeology.

Deploy (historical notes; current flow is release-qud)

CURRENT POLICY (2026-07-06): use release-qud for real Radia deployment. LAB and 100号機 keep radia, cubit-mesh-export, and radia-mcp as editable installs; mdx and hibino consume PyPI wheels. Cubit 2025.12 uses a .ccm backend plus the Cubit-embedded PySide6 toolbar. The old Qt5 .ccl / radia_cubit.ccl path and Cubit 2025.3 symlink recipes in this file are historical only. Do not copy .pyd files by hand, do not revive .ccl, and do not install PySide6 into normal Radia Python. All notebook workbench commands later in this historical file are retired and invalid. Current application QA is Simulink-only; docs notebooks use the separate saved-WebGUI audit.

For current work, read .agents/skills/release-qud/SKILL.md, run tools/release_qud.py, and verify Cubit with cubit-plugin-install --verify-only plus the Cubit smoke tests.

Historical Deploy (2-Stage: LAB editable, 100号機+mdx PyPI)

Philosophy: only the developer (LAB) gets the editable convenience. Every other machine receives the published PyPI wheel so RECORD / entry-points / package-data / cubit-plugin-install regular-file deploy are exercised the same way external users see them. SMB cp shortcuts remain forbidden -- they hide bugs that only surface under a real pip install.

Stage 1  LAB                : pip install -e .                  editable, fastest feedback
                              ↓ the DEVELOPER workflow
Stage 2  100号機 / mdx      : pip install 'radia[cubit]'        PyPI, multi-user distribution
                              ↓ the PUBLIC PyPI test (real-world install)

POLICY (2026-05-02, user-set, REPLACES 2026-05-01 3-tier): 100号機 と mdx は 両方 PyPI install に統一。3-tier (LAB editable / mdx editable / 100号機 PyPI) を 2-tier (LAB editable / 100号機 + mdx PyPI) へ簡素化。

• LAB: pip install -e . (NAS source). developer iteration.
• 100号機: pip install 'radia[cubit]==X.Y.Z' + cubit-plugin-install --all-users. 21 ラボメンバの本番。
• mdx: 100号機 と完全に同じ recipe. PyPI cross-machine consistency probe (release-qud Phase 9). C++ 変更を mdx に届けるには v* tag → CD/PyPI リリースが必要 (LAB の Build.ps1 出力は届かない).

Cubit plugin (100号機 + mdx): 両方とも PyPI wheel から regular-file deploy (cubit-plugin-install --all-users). symlink 体制は使わない.

Application interface policy (2026-07-20): production deploy uses radia[cubit] and one Radia Simulink library. EM, PCB, Motor, and Stream Function are Simulink-only; IH temporarily keeps a notebook comparison. Normal Radia Python must not acquire PySide6. Coreform Cubit's bundled PySide6 remains protected because Cubit owns that embedded runtime.

Retired (2026-05-02): mdx editable + tools/push_pyds_to_mdx.py base64-over-ssh push は通常運用から除外. push_pyds_to_mdx.py は branch-test 用に残置 (PyPI tag を切らずに pre-release .pyd を mdx で試す場合のみ使用).

Each stage uses pip install (editable or from-PyPI). Never copy files to site-packages by hand or via SMB -- if pip install doesn't see it, the real distribution channel won't either.

3 packages in monorepo (ksugahar/Radia), independent PyPI release:

Package	Install (Stage 3)	Purpose
radia	pip install radia	C++ core + Python (FEM/BEM/PEEC, panels)
cubit-mesh-export	pip install cubit-mesh-export	High-order curved mesh export from Cubit
radia-mcp	pip install radia-mcp	MCP servers (radia, ngsolve, cubit, gmsh)

Check Matrix

Each check layer catches a specific class of failure. Upper layers depend on lower layers -- skipping L0/L1 means higher-layer failures become mysterious. All three stages run the full L0-L4 ladder.

Layer	Category	Check	Catches
L0	Syntax	python -m compileall src/radia/	NameError / SyntaxError from recent edits
L0	Install intent	os.path.samefile(src/radia/__init__.py, radia.__file__)	Editable drift to non-editable copy
L0	Wheel manifest	pyproject.toml package-data lists every *.jou/*.step/*.png/*.sol/*.vol/*.msh in src tree	Samples/assets missing from wheel (2026-04-12 calc_mesh_eval.py incident)
L1	File layout	import radia; os.path.isfile(...sample.jou)	Samples / panel_registry.json missing after install
L1	Hash integrity	SHA-256 (CRLF normalised) src vs installed	Content drift (NOT size — LF↔CRLF false-alarms)
L2	Import smoke	import radia.simulink.application, all five DesignSpecs, and the IH notebook adapter	ImportError / circular import in application wiring
L2	Subprocess smoke	python panels/calc_*.py --help for each	argparse rejects flags, missing helper imports
L2	Signature probe	inspect.getsource(f) contains expected token	Edit silently reverted by linter / merge
L3	Simulink QA	simulink-app-health plus focused MATLAB tests	Block/config/DesignSpec/result-artifact drift
L3	Cubit toolbar	cubit-plugin-install --verify-only + toolbar source hash check	Old toolbar/plugin deployed; Cubit still loading stale files
L3	Launcher widget matrix	Static source check: every mode-specific launcher widget has setVisible(ms.needsVol)	"Mesh order: 2" shown for STEP-only mode (2026-04-21 orderRow miss)
L4	End-to-end	Cubit -batch .jou -> export netgen -> .vol	Full pipeline regresses (plugin missing, solver crash)
L4	Application smoke	Trigger each Simulink application through its headless batch path with real inputs -> JSON	block runner + calc script actually solve
L4	Application mode matrix	For EACH active block mode, run its canonical sample E2E on 100号機	New mode shipped without validation (2026-04-21 PEEC-inductance 4.78 nH regression)
L4	Golden-range numeric	Assert L_coil_nH / P_total_W / ... falls in a known range (NOT just exit 0)	Silently-wrong fallback path (e.g. wrong cross-section area → L 100× off)
L4	Multi-user ACL sanity	On 100号機 scan C:\Users\*\AppData\Local\Coreform\Cubit\Coreform\licenses\renewals and any C:\temp\radia_* file; owner MUST match the enclosing user. Admin-owned files under a user AppData are the 2026-04-21 false-path (debug-remote-user SKILL) and block that user's Cubit entirely.	Broken-owner user-cache files from past Admin deploys that shipped but only affect non-admin users

Stage 1 (LAB) runs L0-L3 on LAB only. Stage 2 (100号機 / mdx) runs L0-L4 on each machine after pip install 'radia[cubit]==X.Y.Z' radia-mcp==X.Y.Z cubit-mesh-export==X.Y.Z from PyPI + cubit-plugin-install --all-users. Both machines run the IDENTICAL recipe.

Layer failures cascade: a broken L0 makes L1-L4 results meaningless. Run the layers in order and stop on the first FAIL.

Multi-User Hygiene (100号機, MANDATORY L4)

100号機 is a shared multi-user Windows box with ~20 local lab accounts (keiko / kubota / yano / ...). Any deploy step that writes while running as Administrator must not leave files in another user's profile or in an ACL-restricted shared location. Two bugs have now been observed:

1. 2026-04-21 false-path (skill debug-remote-user): admin copied renewals into each user's AppData\Local\Coreform\Cubit\Coreform\ licenses\ with owner BUILTIN\Administrators. Non-admin users could not read their own cache → RLM -102 every Cubit launch → Cubit silently failed to start → "radia-ngsolve panel crashes" report. Propagated to 8 of 9 non-admin users before detection.

2. 2026-04-24: panel_log.py hard-coded PANEL_LOG_PATH = "C:/radia_panel_log.txt". The root C:\ grants Users:ReadAndExecute only, so the first user to run Cubit (Admin) owned the file and every non-admin subsequent write silently failed. init_panel_log(truncate=True) rotation attempts (rename

   • remove) also failed silently. Non-admin C:\ C++ logs (C:\radia_ccl_log.txt, C:\compact_netgen_debug.log) had the same class of bug — fopen returned NULL for every non-admin Cubit load, so the diagnostic logs were Admin-only by accident.

POLICY — shared-file rules on a multi-user Windows box

Never do this:

• Admin-side cp / Copy-Item / New-Item into another user's %LOCALAPPDATA% or %USERPROFILE%. Owner will be BUILTIN\Administrators and UAC will deny the user read/write.
• Hardcode a shared log / cache path at C:\ root. Root ACL is Users:ReadAndExecute on standard installs.
• Hardcode a shared log in C:\temp\foo.txt. C:\temp has Users:CreateFiles but once Admin creates the file, the file inherits Admin owner, so non-admin append will fail.

Do this instead:

• Per-user filename under C:\temp\: C:\temp\radia_<thing>_<user>.txt. Each user creates their own, owns it, writes it. C:\temp policy is AGENTS.md Temp Directory.
• Per-user %LOCALAPPDATA%: each user's Cubit / Python writes into their own LOCALAPPDATA in their own context. Admin scripts never reach into another user's LOCALAPPDATA — use a logon- trigger Scheduled Task so the work runs as the target user.
• Machine-wide under C:\ProgramData\ (rarely): only when the content is genuinely read-only reference data. Explicit ACL grant Users:ReadAndExecute; never Users:Modify on a shared location.

L4 sanity sweep — before and after every Stage 2 deploy

The same sweep script detects both bug classes above. Run as Administrator on 100号機. deploy is not "done" until this is clean (no BROKEN rows and no non-C:\temp\*_<user>.* hardcoded paths found by grep).

# (1) Broken-owner cache sweep.  Any AppData file whose owner is
#     Administrators / SYSTEM / another user is broken.
cat << 'PS' | ssh 100 'pwsh -ExecutionPolicy Bypass -Command -'
$candidates = @(
  'AppData\Local\Coreform\Cubit\Coreform\licenses\renewals'
)
$bad = @()
Get-ChildItem 'C:\Users' -Directory -EA SilentlyContinue |
  Where-Object { $_.Name -notin @('All Users','Default','Default User','Public') } |
  ForEach-Object {
    foreach ($rel in $candidates) {
      $p = Join-Path $_.FullName $rel
      if (-not (Test-Path $p)) { continue }
      $acl = Get-Acl $p
      $idx = $acl.Owner.LastIndexOf('\')
      $short = if ($idx -ge 0) { $acl.Owner.Substring($idx + 1) } else { $acl.Owner }
      if ($short -ine $_.Name) {
        $bad += [PSCustomObject]@{User=$_.Name;Path=$p;Owner=$acl.Owner}
      }
    }
  }
if ($bad.Count -gt 0) {
  Write-Output "BROKEN-OWNER files detected:"
  $bad | Format-Table -AutoSize
  Write-Output "Fix: takeown /F <path> /A ; icacls <path> /grant Administrators:F ; Remove-Item <path>"
  Write-Output "Then have each affected user log OFF and ON to rebuild with correct owner."
  exit 1
}
Write-Output "OK: all per-user caches have correct owner."
PS

# (2) Hardcoded shared-root path sweep.  Any log / cache file
#     opened at C:\ root (not C:\temp\) is the 2026-04-24 bug class.
ssh 100 'grep -rnE ''"C:\\\\[A-Za-z_][A-Za-z0-9_]*\.(log|txt|json|cfg|ini|cache|tmp)"'' W:\00_CAE\Radia\01_GitHub\src' | grep -v 'C:\\\\temp\\\\'
# expected: no output.  Any hit must be converted to
# C:\temp\<name>_<user>.<ext> (per-user filename in the lab temp dir).

If either sweep finds a hit, fail the deploy loudly — this has silently corrupted a large fraction of users twice in 4 days.

Binary Propagation Policy

POLICY: 全バイナリ (.pyd, .ccm, .ccl) は src/radia/ と packages/cubit-mesh-export/src/cubit_mesh_export/ の 2 箇所 に 存在する必要がある (editable install + wheel + cubit-plugin-install の全てがここから読む)。

Source-of-truth ルール

Phase	Source	Destination(s)
C++ build (Cubit plugin)	cmake --build 出力 dir	src/radia/ + packages/cubit-mesh-export/...
C++ build (radia core)	build-msvc/Release/*.pyd	src/radia/
Python source	src/radia/**/*.py	(no copy — read in place via editable install)
Wheel install (100号機)	pip install -e snapshot	site-packages/radia/
Plugin install (Cubit)	cubit-mesh-export package	Coreform Cubit/bin/{plugins/} via cubit-plugin-install

Build flow (CRITICAL)

1. Build.ps1                                      # MSVC + ninja
     -> build-msvc/Release/_radia_pybind.pyd
     -> src/cubit_plugin/build-pyd/radia_cubit_mesh.pyd
     -> src/cubit_plugin/build-ccm/radia_cubit.ccm
     -> src/cubit_plugin/build-ccm/radia_cubit.ccl   (Qt5 GUI)
2. Build.ps1 が src/radia/ + packages/cubit-mesh-export/ の両方にコピー
3. pip install -e .  +  pip install -e packages/cubit-mesh-export
4. cubit-plugin-install     # cubit-mesh-export -> Cubit/bin/

よくある罠

1. 手動 cmake は別 dir に出力する: cmake --build build-cubit-plugin は build-cubit-plugin/Release/ を作るが Build.ps1 は src/cubit_plugin/build-ccm/ を見ている → コピー先不整合。
2. radia core build がコケると ccm/ccl もコピーされない (Build.ps1 は順番実行)。回避: 修正後の C++ 部分だけ手動で cmake → 出力を src/cubit_plugin/build-ccm/ と src/radia/ の両方に手動コピー。
3. pip install --force-reinstall --no-deps W:\...01_GitHub は source dir を その時点の snapshot で配信する。source dir 内の ファイルが古いと古いものが配信される。先に src/radia/ を 最新化してから pip install すること。

SMB cp は禁止 (2026-04-21 policy)

「pip install を迂回して src/radia/*.py を site-packages に直接 cp -r する」ショートカットは 絶対に使わない。

Why:

• PyPI 配布で end-user が受け取る形態と違う → 実地で動くか判らない
• pip の RECORD / metadata / entry_points が壊れる (旧ファイルが残る)
• editable install が静かに non-editable に上書きされる事故 (PEEC-inductance 2026-04-21)
• wheel に入り損ねる package-data に気付けない

The rule (2026-05-02 updated):

• Stage 1 (LAB): pip install -e . のみ
• Stage 2 (100号機 / mdx): pip install 'radia[cubit]==X.Y.Z' radia-mcp==X.Y.Z cubit-mesh-export==X.Y.Z (PyPI) + cubit-plugin-install --all-users のみ. 両機 完全に同じ recipe.

ショートカットで動いても distribution を検証したことにはならない。 Stage 2 を 2 マシンで二重化することで PyPI 互換性 (wheel manifest / package-data / entry-points / cubit-plugin-install regular-file deploy) を release-qud Phase 9 cross-machine consistency probe で ピン留めする。

radia-mcp Policy

LAB: editable (pip install -e packages/radia-mcp). ローカル レポジトリの packages/radia-mcp/src/radia_mcp/*/server.py を直接参照。 git push で即時反映ではない (LAB の editable は src tree 直読み)。

100号機 / mdx: pip install radia-mcp==X.Y.Z (PyPI). ローカル レポジトリ不可。

cub5_to_vol.py Policy

cub5_to_vol.py (Path B: Python 実装) は Radia リポジトリには配布しない。 netgen fork (ksugahar/netgen) 側で管理・サポートする。 Radia の配布 wheel には bridge.py, cubit_netgen_bridge.py, cub5_to_vol.py は含まれない。

Deployment Strategy per Machine (2026-05-02 simplified)

Component	LAB (ローカル)	100号機 / mdx (PyPI 配布試験)
radia	pip install -e .	pip install 'radia[cubit]==X.Y.Z'
cubit-mesh-export	pip install -e packages/cubit-mesh-export	(radia[cubit] 経由で auto-install)
radia-mcp	pip install -e packages/radia-mcp	pip install radia-mcp==X.Y.Z
PySide6	通常 Radia Python には入れない	Cubit 同梱の PySide6 は保護対象 (削除禁止)
Cubit plugin	(LAB Build.ps1 出力 = src/radia/ に直接)	cubit-plugin-install --all-users (regular-file copy from PyPI wheel)

Simulink migration: [gui] is NOT part of production deploy

Since the application operating surfaces moved to Simulink blocks, production installs should not add the old GUI extra just to get PySide6. Use pip install radia[cubit] for the Cubit/plugin path, verify the five blocks via simulink-app-health, and use ipynb-gui-health only for IH's temporary comparison surface.

Do not uninstall or delete Coreform Cubit's bundled PySide6 under C:/Program Files/Coreform Cubit*/bin/python3/lib/site-packages; Cubit owns that embedded runtime. The no-PySide target applies to the normal Radia Python environment on LAB / 100号機 / mdx / hibino, not to Cubit's private Python.

100号機 / mdx PyPI install の詳細

100号機:

ssh 100 "pip install --upgrade --no-cache-dir 'radia[cubit]==X.Y.Z' radia-mcp==X.Y.Z cubit-mesh-export==X.Y.Z"
ssh 100 'cubit-plugin-install --all-users'
ssh 100 'cubit-plugin-install --verify-only'

mdx (identical recipe):

ssh mdx "pip install --upgrade --no-cache-dir 'radia[cubit]==X.Y.Z' radia-mcp==X.Y.Z cubit-mesh-export==X.Y.Z"
ssh mdx 'cubit-plugin-install --all-users'
ssh mdx 'cubit-plugin-install --verify-only'

新機能をラボメンバに届けるには PyPI release が必要:

1. LAB で fix → 動作確認 → 必要なら別 dev-clone / mdx PyPI で多重確認
2. pyproject.toml の version 上げ → git tag vX.Y.Z
3. git push --tags → CD が PyPI publish
4. PyPI 反映を pip index versions radia で確認 (~5 分)
5. 100号機 で pip install --upgrade 'radia[cubit]==X.Y.Z' ...
6. mdx も同じ recipe (cross-machine consistency probe)
7. kubota 等が 100号機 GUI から動作確認 (Step 8 panel QA)

Remote Access

Target	SSH alias	Default shell	Python	Build chain
LAB (this machine)	ssh lab	bash + pwsh	python	MSVC + Intel oneAPI + NGSolve dev (full)
mdx (163.220.176.228)	ssh mdx	pwsh 7.6	python	binary sync from LAB (no compile)
100号機 (192.168.11.100)	ssh 100	pwsh 7.6	python	none — PyPI wheel only

• All machines use pwsh (PowerShell 7). NOT PowerShell 5.
• SSH config: ~/.ssh/config (aliases 100, mdx, lab).
• File transfer to mdx: scp may be blocked by mdx host policy; use the base64 [IO.File]::WriteAllBytes(...) SSH-stdin pattern instead.
• File transfer to 100号機: SMB (//192.168.11.100/c$/...) or W: drive (LAB ↔ 100号機 share the W:\ NAS mount).
• {} in f-strings conflicts with pwsh ScriptBlock. Use heredoc | ssh host 'pwsh -Command -' and prefer an Out-File to disk + python C:/temp/<script>.py to keep Python source out of the pwsh parser.

Arguments

Argument	Action
(empty / all)	Stage 1 + 2
local	Stage 1 only
100 or 100号機	Stage 2 only
release	git push + tag
mdx or 228	Stage 3 only

---

Stage 1: ローカル ビルド・検証

POLICY: 以下の手順はこの順序で不可分に実行する。途中でスキップや順序変更しない。

cd S:/Radia/01_GitHub

# 1. Kill ALL first (Cubit が .ccl をロック、Python が .pyd をロック)
taskkill //F //IM coreform_cubit.exe
taskkill //F //IM python.exe
# WARNING: SSH-based Stop-Process can only see / kill processes in
# the SSH session, NOT the user's interactive desktop session.  On
# 100号機, a user with Cubit open in their RDP session will NOT be
# killed by this SSH call -- the silent failure looks like success.
# Use WMI for cross-session DETECTION and ask the user to manually
# close Cubit if it is running.
ssh 100 'pwsh -Command "
  $p = Get-CimInstance -ClassName Win32_Process -Filter \"Name = ''coreform_cubit.exe''\"
  if ($p) {
    Write-Host \"WARN: Cubit is running on 100号機 (PID=$($p.ProcessId)).\"
    Write-Host \"      SSH cannot kill cross-session processes.\"
    Write-Host \"      Please close Cubit MANUALLY before continuing the deploy.\"
  } else {
    Write-Host \"Cubit not running on 100号機 -- OK to proceed.\"
  }
"'

# 2. Build (C++ core + ccm + ccl + pyd → src/radia/ AND packages/cubit-mesh-export/)
pwsh -ExecutionPolicy Bypass -File Build.ps1 -Verbose

# 2b. Build the wheel (the artifact that Stage 2 / 3 install)
#     This is the SINGLE SOURCE OF TRUTH for downstream deploy.
#     Same wheel goes to 100号機 (via W:) AND to PyPI (via CI OIDC).
pwsh -ExecutionPolicy Bypass -File Build_Wheel.ps1 -DryRun
# -> dist/radia-X.Y.Z-cp312-cp312-win_amd64.whl
# Also build cubit-mesh-export wheel:
cd packages/cubit-mesh-export
pwsh -ExecutionPolicy Bypass -Command "pip wheel . --no-deps --no-build-isolation --no-cache-dir -w dist"
cd ../..
# -> packages/cubit-mesh-export/dist/cubit_mesh_export-X.Y.Z-*.whl

# 3. Editable install (pip が packages/ 内のバイナリを参照可能にする)
pip install -e . --no-deps
pip install -e packages/cubit-mesh-export --no-deps
# radia-mcp: NOT installed (direct repo reference)

# 4. Deploy plugin (packages/ 内の最新バイナリを Cubit にコピー)
cubit-plugin-install

# 5. Verify (basic import + version)
python -c "import radia; print(radia.__version__)"
python -c "import cubit_mesh_export; print(cubit_mesh_export.__version__)"

# 5b. CRITICAL: verify radia was installed in EDITABLE mode.
#     `pip install .` (non-editable) silently copies src/radia/ into
#     site-packages at install time -- subsequent `src/radia/` edits
#     become INVISIBLE to any code that does `import radia` (including
#     the Cubit .ccl launcher which calls `importlib.util.find_spec`).
#
#     Symptom: deploy looks green, but Cubit Solve menu doesn't show
#     new panel modes you just added (2026-04-21 PEEC-inductance
#     regression). Or panels/calc_*.py adds a relative path to
#     sys.path that ONLY resolves under the source-tree layout, then
#     fails silently in wheel-install layout (2026-05-01 kubota
#     PEEC-BEM `assemble_dense` regression -- bem_sibc_solver lived in
#     examples/ and was unreachable from site-packages/radia/panels/).
#
# HARD GATE (do NOT skip): two independent checks must both pass.
#   (i) `direct_url.json` exists and has `dir_info.editable == true`.
#       This is the canonical PEP 660 marker.  `pip install .` writes
#       `dir_info: {}` (empty) -- non-editable.  `pip install -e .`
#       writes `dir_info: {"editable": true}` -- editable.
#  (ii) `os.path.samefile(src_py, radia.__file__)` is True (defends
#       against an editable .pth that points at a stale clone).
# If EITHER check fails: `pip uninstall -y radia` then
# `pip install -e . --no-deps`.  Do NOT cp files around -- that
# violates Stage 1's editable contract.
python -c "
import radia, os, sys, json
import importlib.metadata as md

src_py = os.path.join(r'S:\Radia\01_GitHub\src\radia', '__init__.py')
sp_py  = radia.__file__
print(f'src: {src_py}')
print(f'sp : {sp_py}')

# (i) PEP 660 marker
try:
    raw = md.distribution('radia').read_text('direct_url.json')
    info = json.loads(raw or '{}')
    editable_flag = bool(info.get('dir_info', {}).get('editable'))
except Exception as e:
    print(f'ERROR: cannot read direct_url.json: {e}')
    sys.exit(1)
print(f'direct_url.json editable = {editable_flag}')
if not editable_flag:
    print('ERROR: radia was installed non-editable (pip install . instead of pip install -e .).')
    print('       Fix: pip uninstall -y radia ; pip install -e . --no-deps')
    sys.exit(1)

# (ii) samefile cross-check
try:
    ok = os.path.samefile(src_py, sp_py)
except OSError:
    ok = False
if not ok:
    print('ERROR: radia.__file__ does not resolve to src/radia/__init__.py.')
    print('       The editable .pth is pointing at a stale clone.')
    print('       Fix: pip uninstall -y radia ; pip install -e . --no-deps')
    sys.exit(1)

print('EDITABLE install OK (PEP 660 marker + samefile both pass)')
sys.exit(0)
"

# 6. Verify samples / GUI assets are reachable from the panel
python -c "
import radia, os
panels = os.path.join(os.path.dirname(radia.__file__), 'panels')
samples = os.path.join(panels, 'samples')
assert os.path.isfile(os.path.join(samples, 'ih_bem_sample.jou')), 'ih_bem_sample.jou missing'
assert os.path.isfile(os.path.join(samples, 'ih_fem_sample.jou')), 'ih_fem_sample.jou missing'
assert os.path.isfile(os.path.join(samples, 'ih_fem_kelvin_sample.jou')), 'ih_fem_kelvin_sample.jou missing'
assert os.path.isfile(os.path.join(samples, 'em_sample.jou')), 'em_sample.jou missing'
assert os.path.isfile(os.path.join(samples, 'pcb_sample.jou')), 'pcb_sample.jou missing'
assert os.path.isfile(os.path.join(panels, 'panel_registry.json')), 'panel_registry.json missing'
print('OK: panels/samples + panel_registry.json reachable from radia package')
"

# 7. Run verify-deploy: confirm Python actually loads the freshly-edited
#    src/radia/**/*.py files (mtime + signature checks). This catches
#    stale .pyc, shadow site-packages installs, and UNC vs S: drive
#    confusion. See `.Codex/skills/verify-deploy/SKILL.md`.
python -c "
import inspect, os
import importlib

# Phase 1: install state
import radia
print(f'  radia.__file__ = {radia.__file__}')

# Phase 2: mtime sanity
for name in ['radia.bem_coupled_solver', 'radia.simulink.application',
             'radia.bem_inductance']:
    m = importlib.import_module(name)
    st = os.stat(m.__file__)
    print(f'  {name:<32} mtime={int(st.st_mtime)} size={st.st_size}')

# Phase 4: signature check (BEM coupled per-DOF f_back)
from radia.bem_coupled_solver import (
    CoupledBEMSolver, assemble_back_reaction_RHS)
src = inspect.getsource(assemble_back_reaction_RHS)
assert 'CoefficientFunction' in src and 'LinearForm' in src, \
    'BEM coupled per-DOF f_back not loaded'
print('  OK: BEM coupled per-DOF f_back is active')

# Phase 6: Simulink application wiring
from radia.simulink.application import APPLICATIONS
assert set(APPLICATIONS) == {'em', 'pcb', 'motor', 'streamfunction', 'ih'}
print('  OK: five Simulink applications')
"

# 7b. L0 — Syntax check all Python files (catches NameError / typo
#     in any radia_*.py / panels/*.py before runtime)
python -m compileall -q src/radia 2>&1 | tail -5
# Exit non-zero + stderr if any file fails.

# 7c. L0 — Wheel manifest audit (WARN; FAIL runs in Wheel Sanity Check)
#     Scope: panels/samples/ only.  The broader src tree has legacy
#     orphan .vol/.msh files (historical mesh debug outputs) that we
#     clean up at wheel-build time via exclude-package-data.
#     This narrow check catches: user adds a new .jou / .step /
#     .sol /.vol to panels/samples/ but forgets to extend the
#     pyproject.toml glob.  The full, repo-wide audit runs in the
#     release-gate "Wheel Sanity Check" section below.
python -c "
import fnmatch, sys, tomllib
from pathlib import Path
root = Path(r'S:\Radia\01_GitHub')
src = root / 'src' / 'radia'
samples = src / 'panels' / 'samples'
cfg = tomllib.loads((root / 'pyproject.toml').read_text(encoding='utf-8'))
pd = cfg.get('tool', {}).get('setuptools', {}).get('package-data', {}).get('radia', [])
excl = cfg.get('tool', {}).get('setuptools', {}).get('exclude-package-data', {}).get('radia', [])

USER_SAMPLES = ('.jou', '.step', '.stp', '.sol', '.py', '.json')
failed = []
for f in samples.rglob('*'):
    if not f.is_file(): continue
    if '__pycache__' in f.parts: continue
    if f.suffix not in USER_SAMPLES: continue
    rel = f.relative_to(src).as_posix()
    if any(fnmatch.fnmatch(rel, pat) for pat in pd): continue
    if any(fnmatch.fnmatch(rel, pat) for pat in excl): continue
    failed.append(rel)
if failed:
    print(f'ERROR: {len(failed)} panel sample file(s) would NOT ship in wheel:')
    for rel in failed[:15]:
        print(f'  {rel}')
    print('  Fix: add a glob to pyproject.toml [tool.setuptools.package-data].radia')
    sys.exit(1)
print(f'OK: all panels/samples/ user-visible files matched by package-data')
"

# 7d. L2 — Import smoke for every radia_*.py (catches ImportError
#     in new panel modules before the user hits "Solve")
python -c "
import importlib, sys
from pathlib import Path
failures = []
for f in sorted((Path('src/radia')).glob('radia_*.py')):
    name = 'radia.' + f.stem
    try:
        importlib.import_module(name)
        print(f'  {name:<40} OK')
    except Exception as e:
        print(f'  {name:<40} FAIL -- {e}')
        failures.append(name)
if failures:
    sys.exit(1)
"

# 7d2. L3 — Launcher widget visibility contract (static source audit)
#      Every QWidget row in the Radia-NGSolve launcher dialog must
#      either be ALWAYS VISIBLE (mode-independent) or gated by
#      `setVisible(ms.needsVol)` / `setVisible(ms.someFlag)`.  A row
#      with no setVisible call is treated as always-visible and may
#      show nonsense for STEP-only modes (e.g. "Mesh order: 2" when
#      no mesh is exported — 2026-04-21 orderRow miss).
python -c "
import re, sys
ccl = open(r'S:\Radia\01_GitHub\src\cubit_plugin\RadiaComp.cpp',
           encoding='utf-8').read()
# Find the Radia-NGSolve launcher function body (between 'Radia-NGSolve'
# QDialog and the end of its updateLabels lambda scope).
m = re.search(r'dlg\.setWindowTitle\(\"Radia-NGSolve\"\)(.*?)dlg\.exec\(\)',
              ccl, re.DOTALL)
if not m:
    print('WARN: could not locate Radia-NGSolve dialog body')
    sys.exit(0)
body = m.group(1)
# Enumerate QWidget* row containers declared in this scope
rows = re.findall(r'QWidget\s*\*\s*(\w+Row|\w+Group)\s*=', body)
# Each must have a setVisible(...) inside updateLabels
updateLabels = re.search(r'updateLabels\s*=\s*\[[^\]]*\]\(\)\s*\{(.*?)\};',
                          body, re.DOTALL)
gated = set(re.findall(r'(\w+)->setVisible', updateLabels.group(1))) if updateLabels else set()
missing = [r for r in rows if r not in gated]
print(f'Launcher row widgets: {rows}')
print(f'Rows gated by updateLabels: {sorted(gated)}')
if missing:
    print(f'WARN: these rows are never hidden (always visible):')
    for r in missing:
        print(f'  {r}')
    print('  If they should hide for STEP-only modes, add setVisible(ms.needsVol).')
else:
    print('OK: every launcher row is explicitly gated by updateLabels')
"

# 7e. L2 — Subprocess argparse smoke: run --help on every calc_*.py.
#     If a calc script raises at import time (missing helper, wrong
#     path), --help errors out instead of printing the flags.
pwsh -Command '
  $panels = "S:\Radia\01_GitHub\src\radia\panels"
  $fails = @()
  foreach ($f in Get-ChildItem "$panels\calc_*.py" -File) {
    $out = & python $f.FullName --help 2>&1
    if ($LASTEXITCODE -ne 0) {
      Write-Host "  $($f.Name) FAIL ($LASTEXITCODE)"
      Write-Host "    $($out | Select-Object -First 3)"
      $fails += $f.Name
    } else {
      Write-Host "  $($f.Name) OK"
    }
  }
  if ($fails.Count -gt 0) { exit 1 }
'

# 7f. L1 — Static probe for non-package imports in panels/calc_*.py.
#     `--help` runs the module top-level only; lazy imports inside
#     solve_*() bodies are NOT exercised.  The 2026-05-01 kubota
#     PEEC-BEM `assemble_dense` regression is the canonical case:
#     calc_peec_bem.py did `from bem_sibc_solver import ...` (bare name)
#     after sys.path.insert(BEM_REF) where BEM_REF was a relative
#     path that resolves under source-tree but NOT under
#     site-packages/radia/panels/ (the wheel-install layout).
#     Subprocess --help did not import bem_sibc_solver (lazy, inside
#     the solver function).  A path-resolution bug shipped silently.
#
#     RULE: every `from <X> import ...` in src/radia/panels/calc_*.py
#     where X is bare (no dot) MUST be one of:
#       (a) a stdlib module
#       (b) a sibling .py in the same panels/ dir
#       (c) a top-level dependency listed in pyproject.toml
#     ANY other bare-name import is rejected -- those rely on a
#     sys.path.insert() of a relative dir, which is the kubota
#     class.  Use `from radia.<name> import ...` instead.
python -c "
import ast, sys, tomllib
from pathlib import Path

repo = Path(r'S:\Radia\01_GitHub')
src_radia = repo / 'src' / 'radia'
panels = src_radia / 'panels'

# panels/calc_common.py adds src/radia/ to sys.path at module load.
# So bare 'from em_material import ...' resolves to radia/em_material.py.
# Whitelist: stdlib + sibling in panels/ + sibling in src/radia/ + dep.
panel_siblings = {p.stem for p in panels.glob('*.py') if p.stem != '__init__'}
radia_siblings = {p.stem for p in src_radia.glob('*.py') if p.stem != '__init__'}

# Stdlib (Python 3.12)
stdlib = set(sys.stdlib_module_names) | {'__future__'}

# pyproject.toml top-level deps
cfg = tomllib.loads((repo / 'pyproject.toml').read_text(encoding='utf-8'))
deps = []
for dep in cfg.get('project', {}).get('dependencies', []):
    deps.append(dep.split()[0].split('>')[0].split('=')[0].split('<')[0].strip())
for grp in cfg.get('project', {}).get('optional-dependencies', {}).values():
    for dep in grp:
        deps.append(dep.split()[0].split('>')[0].split('=')[0].split('<')[0].strip())
deps_top = {d.replace('-', '_').lower() for d in deps if d}
# Common indirect imports we know are present. PySide6 is intentionally absent
# from normal Radia Python after the Simulink interface migration.
deps_top |= {'mcp', 'cubit_mesh_export', 'pyvista', 'matplotlib',
             'scipy', 'pytest'}
deps_top_lower = {x.lower() for x in deps_top}

allowed = stdlib | panel_siblings | radia_siblings | deps_top_lower
fails = []
for f in sorted(panels.glob('calc_*.py')):
    src = f.read_text(encoding='utf-8')
    tree = ast.parse(src, filename=str(f))
    for node in ast.walk(tree):
        if not isinstance(node, ast.ImportFrom):
            continue
        mod = node.module or ''
        # `from X import` -- only top-level X matters here
        top = mod.split('.')[0] if mod else ''
        if not top:
            continue  # relative import 'from . import ...'
        if top.lower() in allowed:
            continue
        if top.startswith('_'):
            continue
        # Resolve path: bare top-level (no dot) AND not whitelisted
        if '.' not in mod:
            fails.append(f'{f.name}:{node.lineno}: from {mod} import ...')

if fails:
    print('ERROR: panels/calc_*.py has non-package bare imports.')
    print('  These work only in source-tree layout; they break in wheel install.')
    print('  Use `from radia.<name> import ...` or move the helper into src/radia/.')
    for line in fails:
        print(f'  {line}')
    sys.exit(1)
print(f'OK: panels/calc_*.py imports all resolve via stdlib / sibling / declared dep')
"

# 7g. L0 — Version consistency: pyproject.toml <-> __init__.py for ALL 3 packages.
#     The 2026-05-02 cubit-mesh-export 0.7.3 release failed at the CI
#     wheel-verify step because pyproject.toml was bumped to 0.7.3 but
#     __init__.py was left at 0.7.2 -- the CI caught it (the publish
#     step did NOT ship the broken wheel), but the developer had to
#     bump to 0.7.4 + retry the whole tag-CI-publish loop (~30 min).
#
#     Worse: the same scan found radia-mcp 0.36.6 had ALREADY shipped to
#     PyPI with a stale __init__.__version__ = "0.34.2" because the
#     release-radia-mcp.yml CI had no version-consistency check at all
#     -- so `python -c "import radia_mcp; print(radia_mcp.__version__)"`
#     on the published wheel returned "0.34.2" while pip shows 0.36.6.
#     This silent mismatch persisted for 1 release cycle before being
#     caught by an audit, and required a 0.36.7 patch release.
#
#     LESSON: don't rely on the per-package release CI to catch this.
#     Run the scan locally in Stage 1 BEFORE pushing the version-bump
#     commit -- catches the bug 0 minutes after introduction, no
#     wasted PyPI tags.
#
#     Scan rule: each package's pyproject.toml [project].version MUST
#     match its top-level __init__.py `__version__ = "..."` literal.
#     Failure here aborts the deploy.
python -c "
import re, sys, tomllib
from pathlib import Path
repo = Path(r'S:\Radia\01_GitHub')
packages = [
    ('radia',
     repo / 'pyproject.toml',
     repo / 'src' / 'radia' / '__init__.py'),
    ('cubit-mesh-export',
     repo / 'packages' / 'cubit-mesh-export' / 'pyproject.toml',
     repo / 'packages' / 'cubit-mesh-export' / 'src'
          / 'cubit_mesh_export' / '__init__.py'),
    ('radia-mcp',
     repo / 'packages' / 'radia-mcp' / 'pyproject.toml',
     repo / 'packages' / 'radia-mcp' / 'src' / 'radia_mcp' / '__init__.py'),
]
fails = []
for name, pyproject_path, init_path in packages:
    if not pyproject_path.is_file():
        fails.append(f'{name}: pyproject.toml missing at {pyproject_path}')
        continue
    cfg = tomllib.loads(pyproject_path.read_text(encoding='utf-8'))
    pyproject_ver = cfg.get('project', {}).get('version')
    if not pyproject_ver:
        fails.append(f'{name}: no [project].version in pyproject.toml')
        continue
    if not init_path.is_file():
        fails.append(f'{name}: __init__.py missing at {init_path}')
        continue
    init_text = init_path.read_text(encoding='utf-8')
    m = re.search(r'^__version__\s*=\s*[\"\\\']([^\"\\\']+)', init_text, re.M)
    if not m:
        fails.append(f'{name}: no __version__ literal in __init__.py')
        continue
    init_ver = m.group(1)
    if pyproject_ver != init_ver:
        fails.append(
            f'{name}: pyproject.toml=\"{pyproject_ver}\" '
            f'vs __init__.py=\"{init_ver}\" '
            f'-- bump both, not just one')
    else:
        print(f'  OK {name:25s} {pyproject_ver}')
if fails:
    print()
    print('ERROR: pyproject.toml <-> __init__.py version mismatch:')
    for f in fails:
        print(f'  {f}')
    print()
    print('Fix: bump BOTH the pyproject.toml [project].version AND the')
    print('     __init__.__version__ literal in lockstep, then re-run.')
    sys.exit(1)
print('OK: all 3 packages have consistent pyproject.toml / __init__.py versions')
"

順序を守る理由:

• 1 の前にビルドすると .pyd/.ccl がロックされてコピー失敗
• 3 の前に 4 を実行すると古いバイナリがデプロイされる
• 6 の verify を省略すると wheel install で .jou 不在に気付かない (panel が起動しても sample 一覧が空になる)
• 7b-g の L0/L1/L2 チェックは コードレビューより先に走る (static → run-time の順)
• 個別実行は禁止。必ず 1→2→3→4→5→5b→6→7→7b→7c→7d→7e→7f→7g→8 の順序

Step 8 - Simulink Application QA (MANDATORY when interface code changed)

POLICY: src/radia/simulink/application.py, matlab/+radia/+simulink/, src/radia/*_design.py, src/radia/panels/calc_*.py, or application_interface_manifest.json を編集したら、 simulink-app-health を production gate として使う。公開docs notebookを 編集した場合は ipynb-gui-health も実行する。normal Radia Python に PySide6 を追加して旧 desktop panel や notebook workbench を検証対象に戻してはいけない。

自動チェック:

python -m pytest tests/test_simulink_application.py -q
python -m pytest tests/test_application_interface_manifest.py -q
python -m pytest packages/radia-mcp/tests/test_document_meta_notebook_audit.py -q
python tools/audit_new_panel_contract.py

このゲートは次を同時に確認する:

• 5つの application block と versioned config / DesignSpec が一致している
• explicit trigger 1回につき headless solve が1回だけ走る
• success / failure の双方で durable result.json が残る
• application notebook workbench が復活していない
• docs examples の保存済み WebGUI と JSON SHA が一致している

旧PySide panel QA は legacy adapter を直接修正した時の補助チェックに限る。 通常のリリース判定は Simulink application blocks と cubit-plugin-install --verify-only / cubit-smoke-test で行う。

Panel Samples Quality Gate (Stage 1, MANDATORY)

POLICY (2026-04-21): src/radia/panels/samples/ は wheel に入って end-user の Browse 初期値になる。徹底的にデバッグした高品質なもの だけ を置くこと。新しい sample を panels/samples/ に追加する場合、 Stage 1 で以下を通すまで deploy しない:

Step	Action
1	Cubit で play <sample>.jou が最後まで通る
2	必要なら export netgen/.../gmsh が正しい .vol / .msh を吐く
3	対応するパネルの Run で期待値が出る (LAB 上で subprocess 実行)
4	tests/panels/test_<feature>_golden.py が PASS (数値を lock)
5	pyproject.toml の package-data に glob が入っている (wheel manifest audit 7c で確認)

WIP / 研究用 sample は examples/ か tests/**/fixtures/ に置く。 panels/samples/ に置くのは release action であって dev convenience ではない。

Why: end-user は「壊れた sample」と「user error」を区別できない。 試して失敗 → 「このパネル使えない」という判断になり、壊れた sample 1 つで パネル全体の信頼が失われる。PEEC-inductance 2026-04-21 で 3turncoil.stp が walker-hang 状態のまま samples に入っていて Kubota が 初回に遭遇した教訓。

Wheel に含まれる GUI アセット (samples + registry)

Build.ps1 はバイナリだけでなく GUI アセット も wheel に詰める。 これらは pyproject.toml の [tool.setuptools.package-data] で 明示的に列挙する必要がある (setuptools は *.py 以外を自動収集しない):

[tool.setuptools.package-data]
radia = [
    "*.pyd", "*.ccm", "*.ccl", "*.png", "*.ico",
    "panels/samples/*.jou",      # IH/EM/PCB の例題ジャーナル
    "panels/panel_registry.json", # パネル定義 (radia_*.py が参照)
    "CONVENTIONS.md",
]

新しい例題 .jou を追加したら、上記リストを更新せずに wheel を 作っても wheel には入らない。pip install 後の Cubit パネルで sample が選べなくなるので必ず確認すること。

確認コマンド (wheel build 後):

python -c "
import zipfile
z = zipfile.ZipFile('dist/radia-X.Y.Z-cp312-cp312-win_amd64.whl')
for n in sorted(z.namelist()):
    if 'samples' in n or 'panel_registry' in n:
        print(' ', n)
"

Stage 2: 100号機 + mdx (PyPI install)

POLICY (2026-05-02): The 3-tier model (LAB editable, mdx editable, 100号機 PyPI) collapsed to 2-tier: LAB editable + 100号機/mdx PyPI. mdx editable is retired (2026-05-01 → 2026-05-02 lifespan). Both 100号機 and mdx now run the IDENTICAL PyPI install recipe + Cubit plugin regular-file deploy. Verified on mdx 2026-05-02: cubit-plugin-install --verify-only 5/5 [OK] sha256 match, Phase C-heavy fixture L_coil = 138.159 nH.

Stage 2 — 100号機 / mdx PyPI install (COMPLETE RECIPE 2026-05-02)

Run AFTER pip index versions radia confirms the new wheel is live. Both 100号機 and mdx use exactly this recipe. Substitute the SSH target (192.168.11.100 for 100号機, mdx for mdx). Do not add the old GUI extra for PySide6; Simulink blocks are the canonical application surface, with the IH notebook retained only for comparison.

TARGET=192.168.11.100   # or 'mdx'
VERSION=4.24.0          # version being released

cat << PS | ssh "$TARGET" 'pwsh -ExecutionPolicy Bypass -Command -'
\$ErrorActionPreference = 'Stop'

# Step 1: stop locks (python.exe / mcp-server-* / Cubit hold .pyd open).
Get-Process -ErrorAction SilentlyContinue | Where-Object {
  \$_.Name -like 'mcp-server*' -or \$_.ProcessName -eq 'coreform_cubit' -or \$_.ProcessName -eq 'cubit'
} | ForEach-Object { Stop-Process -Id \$_.Id -Force }
Start-Sleep -Seconds 2

# Step 2: tear down any prior install state.  uninstall first so
# editable-from-NAS or editable-from-clone leftovers do not shadow the
# PyPI wheel.
pip uninstall -y radia radia-mcp cubit-mesh-export 2>&1 | Select-String "Successfully|already|not installed"

# Step 3: PyPI install.  The cubit extra pulls cubit-mesh-export.
# Notebook panel QA is handled without installing PySide6 into normal Python.
pip install --no-cache-dir 'radia[cubit]==$VERSION' 'radia-mcp==$VERSION' 'cubit-mesh-export==$VERSION'

# Step 4: deploy Cubit plugin as regular files (NOT symlinks).
cubit-plugin-install --all-users
cubit-plugin-install --verify-only
PS

The --verify-only output should show:

[OK] radia_cubit.ccm    (NNNNNN bytes, sha256 match)
[OK] radia_cubit.ccl    (NNNNNN bytes, sha256 match)
[OK] radia_cubit_mesh.cp312-win_amd64.pyd  (NNNNNN bytes, sha256 match)
[OK] nglib.dll          (NNNNNN bytes, sha256 match)
[OK] ngcore.dll         (NNNNNN bytes, sha256 match)
compat: radia X.Y.Z <-> cubit-mesh-export A.B.C compatible
[OK] every expected binary present and matches package source.

End-to-end smoke (run on the same SSH target after install):

ssh "$TARGET" 'cubit-smoke-test'      # Cubit batch + export netgen + .vol parse
ssh "$TARGET" 'python -m pytest tests/test_simulink_application.py tests/test_application_interface_manifest.py -q'

Both probes are mandatory before declaring the deploy complete. On a source checkout, also run the docs notebook audit when notebooks changed. See also release-qud Phase 8c (verify) and Phase 8d (smoke).

Stage 2 — LEGACY (mdx editable, 2026-05-01 only) -- RETIRED 2026-05-02

Critical points -- if any of these is missed, the editable install silently breaks (the 2026-05-01 incident took 3 hours to diagnose):

1. mdx has NO MSVC, NO Intel MKL, NO gh CLI. C++ binaries must come from LAB via base64-over-ssh push, NOT from download_binaries.sh (gh-CLI-dependent).
2. cubit_mesh_export legacy shadow at C:\Program Files\Python312\Lib\site-packages\cubit_mesh_export.py (single-file legacy module) MUST be deleted before editable install -- pip uninstall does NOT touch it because it was placed outside pip metadata, and the shadow takes precedence over the editable package's __init__.py.
3. Three packages, not two: radia, cubit-mesh-export, and radia-mcp. The earlier recipe missed radia-mcp.
4. radia_cubit_mesh.pyd lives in TWO places on the local clone -- src/radia/ AND packages/cubit-mesh-export/src/cubit_mesh_export/ -- and BOTH must be populated. LAB Build.ps1 writes both; on mdx (no build) we push both via base64-over-ssh.
5. After the recipe, run the full verification probe which checks editable=True for all 3 packages, the C++ kernel imports, and the radia_cubit_mesh extension imports.

Step A — first-time setup OR full recovery on mdx

Run this from LAB. Idempotent: re-runnable any time mdx drifts.

# A1. Fix mdx-side shadows + install editable.
cat << 'PS' | ssh mdx 'pwsh -ExecutionPolicy Bypass -Command -'
# Stop locks
Get-Process -ErrorAction SilentlyContinue | Where-Object { $_.Name -like 'mcp-server*' } | ForEach-Object { Stop-Process -Id $_.Id -Force }
Start-Sleep -Seconds 2

# Delete legacy single-file shadow at site-packages (NOT touched by pip uninstall).
$shadow = "C:\Program Files\Python312\Lib\site-packages\cubit_mesh_export.py"
$shadowPyc = "C:\Program Files\Python312\Lib\site-packages\__pycache__\cubit_mesh_export.cpython-312.pyc"
if (Test-Path $shadow) { Remove-Item $shadow -Force; Write-Host "removed legacy shadow" }
if (Test-Path $shadowPyc) { Remove-Item $shadowPyc -Force; Write-Host "removed shadow .pyc" }

# Ensure local clone exists
$dest = 'C:\Radia\01_GitHub'
if (-not (Test-Path $dest)) {
    New-Item -ItemType Directory 'C:\Radia' -Force | Out-Null
    git clone https://github.com/ksugahar/Radia.git $dest
}
Set-Location $dest

# Uninstall any stale (editable or PyPI)
pip uninstall -y radia cubit-mesh-export radia-mcp 2>&1 | Select-String "Successfully|already|not installed"

# Install editable: ALL THREE packages
pip install --no-deps --no-cache-dir -e . 2>&1 | Select-String "Successfully"
pip install --no-deps --no-cache-dir -e packages/cubit-mesh-export 2>&1 | Select-String "Successfully"
pip install --no-deps --no-cache-dir -e packages/radia-mcp 2>&1 | Select-String "Successfully"
PS

Step B — push C++ binaries from LAB to mdx

Run this from LAB after every Build.ps1 whose output mdx should see.

python tools/push_pyds_to_mdx.py

The tool:

• Stops any python.exe / pythonw.exe / jupyter* / mcp-server-* on mdx first (otherwise WriteAllBytes fails with WinError 32: process cannot access the file because it is being [used by another process] -- Python keeps the loaded .pyd open).
• Pushes every .pyd to its expected location(s). Note that radia_cubit_mesh.pyd is mirrored to TWO paths -- src/radia/ AND packages/cubit-mesh-export/src/cubit_mesh_export/. The src/radia/ copy is for download_binaries.sh parity (in case mdx ever gains the gh CLI); the packages/.../cubit_mesh_export/ copy is what cubit-mesh-export editable install expects at import time.

Tool source: tools/push_pyds_to_mdx.py. The tool is destructive (terminates in-flight Python sessions on mdx); mdx is a verification target only so this is acceptable.

Step C — day-to-day pull (after LAB git push)

ssh mdx 'git -C C:\Radia\01_GitHub fetch origin && git -C C:\Radia\01_GitHub checkout main && git -C C:\Radia\01_GitHub pull --ff-only'
# Then re-push C++ pyds (Step B) if Build.ps1 ran on LAB.

Step D — full editable verification

cat << 'PY' | ssh mdx 'python -'
import importlib.metadata as md
import os, json, math, sys

# 1. radia editable
import radia
expected_src = r"C:\Radia\01_GitHub\src\radia\__init__.py"
assert os.path.samefile(radia.__file__, expected_src), f"radia not from local clone: {radia.__file__}"
info = json.loads(md.distribution("radia").read_text("direct_url.json") or "{}")
assert info.get("dir_info", {}).get("editable"), "radia not editable"
print(f"[OK] radia v{radia.__version__} editable from {radia.__file__}")

# 2. _radia_pybind C++ kernel imports
from radia._radia_pybind import _average_demag_tensor, _AssembleSLDL_Galerkin
import numpy as np
A = _average_demag_tensor(np.array([0.,0.,0.]), np.array([1.,1.,1.]),
                          np.array([0.,0.,0.]), np.array([1.,1.,1.]))
assert abs(A.trace() + 1.0) < 1e-13, f"self-cube trace wrong: {A.trace()}"
print(f"[OK] _radia_pybind C++ kernel works (self-cube trace={A.trace():.4f})")

# 3. radia-mcp editable
print(f"[OK] radia-mcp v{md.version('radia-mcp')}")

# 4. cubit-mesh-export editable + C++ extension
import cubit_mesh_export
expected_cme = r"C:\Radia\01_GitHub\packages\cubit-mesh-export\src\cubit_mesh_export\__init__.py"
assert os.path.samefile(cubit_mesh_export.__file__, expected_cme), \
    f"cubit_mesh_export NOT from local clone (legacy shadow?): {cubit_mesh_export.__file__}"
from cubit_mesh_export import check, radia_cubit_mesh
print(f"[OK] cubit-mesh-export v{cubit_mesh_export.__version__} editable + radia_cubit_mesh.pyd loads")

# 5. v4.23.0 IH formula (regression check on the C++ binding)
from radia.analytical_formulas import cylinder_axial_eddy_loss
P = cylinder_axial_eddy_loss(1000.0, 0.02, 2*math.pi*10000, 5.8e7)
expected_P = 1.6121e+00
assert abs(P - expected_P) / expected_P < 0.01, f"IH formula drift: {P} vs {expected_P}"
print(f"[OK] IH cylinder Cu 20mm 10kHz = {P:.4e} W/m (expected ~1.6121)")

print("=== mdx editable verification ALL OK ===")
PY

Common failure modes (from 2026-05-01/02 incidents)

Symptom	Root cause	Fix
ImportError: cannot import name 'check' from 'cubit_mesh_export'	legacy cubit_mesh_export.py shadow at site-packages	Step A removes it
AttributeError: module 'cubit_mesh_export' has no attribute '__version__'	same legacy shadow	Step A removes it
ImportError: DLL load failed: ... radia_cubit_mesh.pyd	.pyd only in src/radia/ not in packages/.../cubit_mesh_export/ (or vice versa)	Step B pushes BOTH
radia.__version__ works but from radia._radia_pybind import ... fails	_radia_pybind.pyd missing or stale	Step B re-pushes
from radia._radia_pybind import _AssembleSLDL_Galerkin ImportError	mdx pyd predates a recent C++ change on LAB	Step B re-pushes after Build.ps1
download_binaries.sh reports success but no .pyds appear	mdx has no gh CLI; the script silently exits with no error	Don't use it on mdx -- use Step B instead

---

100号機 の drive layout (legacy reference — Stage 2 was 100号機 until 2026-05-01)

100号機 の中で見たときの drive 状況:

Drive	Exists	Notes
S:\	NO	LAB-only. 100号機からは見えない. 使うな
W:\	YES	100号機ローカルドライブ. repo は W:\00_CAE\Radia\01_GitHub
C:\	YES	OS / Python / Coreform Cubit

LAB から 100号機の C: にアクセスするときは SMB: //192.168.11.100/c$/...

ssh 100 'Test-Path S:\'              # -> False (S: は LAB-only)
ssh 100 'Test-Path W:\00_CAE\...'    # -> True

LAB 上で 100号機の repo を直接見たい場合: SMB share //192.168.11.100/work/00_CAE/Radia/01_GitHub (= 100号機の W:\00_CAE\Radia\01_GitHub).

Stage 2 — 100号機: editable + Cubit plugin symlinks (verify only, no install per deploy)

POLICY (2026-04-24, user-set): 100号機 は 修正をすぐに反映する editable にしておく。一回 first-time editable install + Cubit plugin symlink を貼ったあとは、LAB の編集 (Python) と LAB の Build.ps1 出力 (.ccm / .ccl / .pyd) が即 100号機 で live になる。 deploy ごとに pip install も cubit-plugin-install も走らせない。 かなり開発が進んだときだけ PyPI に上げて、その配布形態の 検証は mdx (Stage 3) で行う。100号機 は PyPI 配布の試験対象では ない。

This is the all-users multi-user test — the editable install lives in the system-wide C:\Program Files\Python312\Lib\site-packages\ (direct_url.json points at W:\…\src\radia) which every local user on 100号機 reads. SMB cp shortcuts are still forbidden (see the "SMB cp は禁止" section above).

(a) Pre-flight: FORCE-CLOSE all Cubit sessions

POLICY: Deploy takes priority. Notify each Cubit user with a 30-second warning popup, then force-kill any remaining coreform_cubit.exe via Stop-Process -Id <pid> -Force (cross- session kill-by-PID works with admin rights — it's -Name that's session-scoped). Do NOT wait for user cooperation — .ccl must be replaceable or the deploy is worthless.

User data loss risk: users lose unsaved Cubit state. This is acceptable because the skill emits the warning popup first, and the lab agreement is that deploy is a critical operation.

cat << 'PS' | ssh 100 'pwsh -Command -'
$p = @(Get-CimInstance Win32_Process -Filter "Name='coreform_cubit.exe'" -EA SilentlyContinue)
if ($p.Count -eq 0) {
  Write-Host "Cubit not running -- safe to deploy"
  exit 0
}
Write-Host "Cubit is running; force-close procedure:"
$sessions = @{}
foreach ($proc in $p) {
  $owner = ($proc.GetOwner()).User
  $sess  = $proc.SessionId
  $sessions[$sess] = $owner
  Write-Host ("  PID=$($proc.ProcessId)  User=$owner  Session=$sess")
}
# 1) warn every affected user with a popup (30-s countdown)
foreach ($kv in $sessions.GetEnumerator()) {
  try {
    msg $kv.Key /TIME:30 "[Radia deploy] Cubit will be force-closed in 30 seconds. SAVE YOUR WORK NOW."
    Write-Host "  notified session $($kv.Key) ($($kv.Value))"
  } catch {
    Write-Host "  notify FAILED for session $($kv.Key) -- msg may be disabled"
  }
}
# 2) wait 30 s for voluntary close
Start-Sleep -Seconds 30
# 3) force-kill anything still running (kill-by-PID works cross-session)
$still = @(Get-CimInstance Win32_Process -Filter "Name='coreform_cubit.exe'" -EA SilentlyContinue)
foreach ($proc in $still) {
  $pid_ = $proc.ProcessId
  try {
    Stop-Process -Id $pid_ -Force -EA Stop
    Write-Host "  force-killed PID $pid_"
  } catch {
    # Fallback via WMI Terminate (works when Stop-Process is blocked)
    $proc.Terminate() | Out-Null
    Write-Host "  force-killed PID $pid_ (WMI Terminate fallback)"
  }
}
# 4) verify all gone
Start-Sleep -Seconds 2
$left = @(Get-CimInstance Win32_Process -Filter "Name='coreform_cubit.exe'" -EA SilentlyContinue)
if ($left.Count -gt 0) {
  Write-Host "ERROR: $($left.Count) Cubit process(es) survived force-kill"
  exit 1
}
Write-Host "OK: all Cubit sessions closed, proceeding with deploy"
PS

If the user wants to avoid losing work, they see the 30-second popup and Ctrl+S / File → Save in time. After deploy, re-notify:

ssh 100 'pwsh -Command "
  query session | Select-String \"Active\" | ForEach-Object {
    \$sid = ((\$_.Line -split \"\s+\") | Where-Object { \$_ -match \"^\d+\$\" })[0]
    if (\$sid) { msg \$sid /TIME:30 \"[Radia deploy] Complete. Safe to reopen Cubit.\" }
  }
"'

(b) Editable install + Cubit plugin symlinks (no pip install / no cubit-plugin-install)

POLICY (2026-04-24): 100号機 is editable + symlink. No wheel install, no cubit-plugin-install. LAB Build.ps1 output (in src/radia/ and packages/cubit-mesh-export/...) becomes effective on 100号機 immediately because:

• pip install -e W:\00_CAE\Radia\01_GitHub (radia) and pip install -e W:\00_CAE\Radia\01_GitHub\packages\cubit-mesh-export (cubit-mesh-export) are one-time setup in machine-wide C:\Program Files\Python312\Lib\site-packages\. import radia resolves directly to W:\…\src\radia\__init__.py.
• <Cubit>\bin\radia_cubit.ccl / <Cubit>\bin\plugins\radia_cubit.ccm / <Cubit>\bin\plugins\radia_cubit_mesh.cp312-win_amd64.pyd are SymbolicLinks to the corresponding files on W:. Cubit loads the live LAB build through them.

DO NOT run cubit-plugin-install on 100号機 — the installer copies regular files over the symlinks and the editable-deploy chain breaks. PyPI-style regular-file deploy is exercised on mdx (Stage 3), not 100号機.

DO NOT run pip install (wheel or source) on 100号機 unless one of the following first-time / repair conditions applies:

Condition	Action
Brand-new 100号機 install	Run "First-time editable install" below.
pyproject.toml deps changed (NGSolve / MKL bump)	pip install -e W:\…\01_GitHub --upgrade
Editable drifted (samefile check fails in Step (c))	pip uninstall -y radia cubit-mesh-export; pip install -e W:\…
<Cubit>\bin\…\radia_cubit* is a regular file (not SymbolicLink)	Re-create the 3 symlinks (see below).

Otherwise the only Stage 2 work per deploy is verification that the editable install + symlinks are still healthy.

First-time editable install (one-time setup, admin only)

cat << 'PS' | ssh 100 'pwsh -ExecutionPolicy Bypass -Command -'
# 1. Editable install for ALL local users (machine-wide site-packages)
pip install --no-deps -e "W:\00_CAE\Radia\01_GitHub"
pip install --no-deps -e "W:\00_CAE\Radia\01_GitHub\packages\cubit-mesh-export"

# 2. Cubit plugin symlinks (admin token required for mklink)
$cubit = "C:\Program Files\Coreform Cubit 2025.3"
$repo  = "W:\00_CAE\Radia\01_GitHub"
# Remove any pre-existing regular file before mklink
Remove-Item -Force -EA SilentlyContinue `
    "$cubit\bin\radia_cubit.ccl", `
    "$cubit\bin\plugins\radia_cubit.ccm", `
    "$cubit\bin\plugins\radia_cubit_mesh.cp312-win_amd64.pyd"
# mklink (use cmd /c — pwsh New-Item -ItemType SymbolicLink to UNC paths is fragile)
cmd /c mklink "$cubit\bin\radia_cubit.ccl"                           "$repo\src\radia\radia_cubit.ccl"
cmd /c mklink "$cubit\bin\plugins\radia_cubit.ccm"                   "$repo\src\radia\radia_cubit.ccm"
cmd /c mklink "$cubit\bin\plugins\radia_cubit_mesh.cp312-win_amd64.pyd" `
              "$repo\packages\cubit-mesh-export\src\cubit_mesh_export\radia_cubit_mesh.pyd"
PS

ACL note: W:\00_CAE\Radia\01_GitHub already has NT AUTHORITY\Authenticated Users Modify, so all 21 local users can read the editable install + symlink targets. Only Admin can do mklink, so symlink maintenance always runs as Administrator via SSH.

Per-deploy verification (the normal path)

For an ordinary LAB build → 100号機 propagation, no install or plugin-deploy command runs. Just verify the editable + symlink chain is intact, then go to Step (c) and beyond. The non-destructive preflight is cubit-plugin-install --check-only (introduced in cubit-mesh-export v0.5.2 — does NOT touch any files):

cat << 'PS' | ssh 100 'pwsh -Command -'
"=== Cubit plugin: must be SymbolicLink ==="
$paths = @(
  "C:\Program Files\Coreform Cubit 2025.3\bin\radia_cubit.ccl",
  "C:\Program Files\Coreform Cubit 2025.3\bin\plugins\radia_cubit.ccm",
  "C:\Program Files\Coreform Cubit 2025.3\bin\plugins\radia_cubit_mesh.cp312-win_amd64.pyd"
)
$bad = @()
foreach ($p in $paths) {
  $f = Get-Item $p -EA SilentlyContinue
  if (-not $f) { $bad += "MISSING: $p"; continue }
  if ($f.LinkType -ne "SymbolicLink") { $bad += "REGULAR FILE (expected SymbolicLink): $p"; continue }
  $real = Get-Item $f.Target -EA SilentlyContinue
  if (-not $real) { $bad += "BROKEN LINK: $p -> $($f.Target)"; continue }
  Write-Host ("  OK  {0}" -f $f.Name)
  Write-Host ("      -> {0} ({1} bytes, {2})" -f $f.Target, $real.Length, $real.LastWriteTime)
}
if ($bad) { Write-Host "FAIL:"; $bad | ForEach-Object { Write-Host "  $_" }; exit 1 }

"=== Editable install: import radia must resolve to W:\..\src\radia ==="
python -c "import radia, os; assert os.path.samefile(radia.__file__, r'W:\00_CAE\Radia\01_GitHub\src\radia\__init__.py'); print('  OK editable: ' + radia.__file__)"
python -c "import cubit_mesh_export, os; assert os.path.samefile(cubit_mesh_export.__file__, r'W:\00_CAE\Radia\01_GitHub\packages\cubit-mesh-export\src\cubit_mesh_export\__init__.py'); print('  OK editable: ' + cubit_mesh_export.__file__)"

"=== cubit-plugin-install --check-only (NON-DESTRUCTIVE preflight) ==="
cubit-plugin-install --check-only
PS

Expected output: every symlink resolves, both samefile() asserts pass, preflight prints [OK] preflight clean. Any FAIL means the editable + symlink chain has drifted — repair per the table above.

(c) Verify: editable-install samefile check (NOT hash, NOT size)

Stage 2 is editable, so radia.__file__ and the source file in W:\…\src\radia\__init__.py MUST resolve to the same inode (samefile is True). This is stronger than a hash compare — there is no second copy of the file at all.

If samefile fails, the editable install was silently overwritten by a non-editable pip install (the 2026-04-21 PEEC-inductance regression class, and the 2026-05-01 kubota PEEC-BEM assemble_dense regression where the wheel install put panels/calc_peec_bem.py into site- packages with a BEM_REF relative path that resolved to a non- existent location, masking the real version of bem_sibc_solver). Repair: pip uninstall -y radia cubit-mesh-export then re-run the first-time editable install in Step (b).

The HARD GATE is two-pronged: (i) PEP 660 marker direct_url.json::dir_info.editable == true, AND (ii) samefile() between import and source. Either failing means a pip install (non--e) silently shipped a frozen copy.

cat << 'PY' | ssh 100 'python -'
import os, sys, json
import radia, cubit_mesh_export
import importlib.metadata as md
src_root = r"W:\00_CAE\Radia\01_GitHub\src\radia"
cme_root = r"W:\00_CAE\Radia\01_GitHub\packages\cubit-mesh-export\src\cubit_mesh_export"

print(f"radia v{radia.__version__}  loc={radia.__file__}")
print(f"cubit_mesh_export v{cubit_mesh_export.__version__}  loc={cubit_mesh_export.__file__}")

fails = []
def check_editable(pkg_name):
    """PEP 660 marker check: direct_url.json must say editable=True."""
    try:
        raw = md.distribution(pkg_name).read_text('direct_url.json')
        info = json.loads(raw or '{}')
        flag = bool(info.get('dir_info', {}).get('editable'))
    except Exception as e:
        print(f"  {pkg_name:<48} [NO direct_url.json: {e}]")
        fails.append(pkg_name + " editable marker")
        return
    tag = "OK" if flag else "DRIFT (non-editable)"
    print(f"  {pkg_name:<48} editable={flag} [{tag}]")
    if not flag:
        fails.append(pkg_name + " editable marker")

def check(import_path, src_path, label):
    try:
        ok = os.path.samefile(import_path, src_path)
    except OSError:
        ok = False
    tag = "OK" if ok else "DRIFT"
    print(f"  {label:<48} [{tag}]")
    if not ok:
        print(f"    import: {import_path}")
        print(f"    src   : {src_path}")
        fails.append(label)

check_editable("radia")
check_editable("cubit-mesh-export")
check(radia.__file__, os.path.join(src_root, "__init__.py"), "radia.__init__ is editable from W:")
check(cubit_mesh_export.__file__, os.path.join(cme_root, "__init__.py"), "cubit_mesh_export.__init__ is editable from W:")

# Spot-check a few panel files: if these aren't samefile, something
# is shadowing the editable install (rogue site-packages copy, conda env, etc.)
for rel in ["radia_ih.py", "radia_gui_base.py", "radia_peec_inductance.py",
            "panels/calc_peec_inductance.py", "panels/panel_registry.json"]:
    src = os.path.join(src_root, rel.replace("/", os.sep))
    # Compute the would-be import path: editable means same path as src
    check(src, src, f"src present: {rel}") if not os.path.isfile(src) else None
    if os.path.isfile(src):
        print(f"  src present: {rel:<40} ({os.path.getsize(src)} bytes)")

sys.exit(0 if not fails else 1)
PY

Any DRIFT line means pip install -e was silently overwritten or shadowed. Do NOT proceed — fix per the table in Step (b).

(d) End-to-end verify (application runner and docs contract)

cat << 'PS' | ssh 100 'pwsh -Command -'
Set-Location 'W:\00_CAE\Radia\01_GitHub'
python -m pytest tests/test_simulink_application.py -q
python -m pytest tests/test_application_interface_manifest.py -q
python -m pytest packages/radia-mcp/tests/test_document_meta_notebook_audit.py -q
python tools/audit_new_panel_contract.py
PS

Expected: Simulink application and docs notebook contracts PASS on 100号機.

(e) Post-deploy verification: "really deployed" HARD GATE

Worst failure mode: deploy looks green but isn't actually in effect. Old .ccl in a zombie Cubit process, wrong Python on PATH, stale .pyc cache, pip install silently skipped. These produce silent wrong state that propagates into the next bug hunt.

Run these checks AFTER pip install + cubit-plugin-install. Every check FAILs LOUDly — no warnings-treated-as-OK.

cat << 'PS' | ssh 100 'pwsh -Command -'
$fail = @()

# --- Check 1: no zombie Cubit older than the .ccl ---
# Any Cubit process running NOW must have started AFTER the .ccl's
# LastWriteTime.  If one predates, it's running the old plugin.
$ccl = Get-Item "C:\Program Files\Coreform Cubit 2025.3\bin\radia_cubit.ccl"
$cubs = @(Get-CimInstance Win32_Process -Filter "Name='coreform_cubit.exe'" -EA SilentlyContinue)
foreach ($c in $cubs) {
  $t = ([Management.ManagementDateTimeConverter]::ToDateTime($c.CreationDate))
  if ($t -lt $ccl.LastWriteTime) {
    $fail += "ZOMBIE: PID $($c.ProcessId) started $t (before .ccl update at $($ccl.LastWriteTime))"
  }
}

# --- Check 2: .ccl on disk contains NEW strings ---
# grep for strings known to be in the latest .ccl build.  If a string
# is missing, either the build was skipped or Cubit/bin has stale .ccl.
$cclText = Get-Content -LiteralPath $ccl.FullName -Raw -Encoding Byte | ForEach-Object {
  [System.Text.Encoding]::ASCII.GetString($_)
}
# (ScriptBlock-safe: PowerShell joins byte output into one string)
$cclBin = [System.IO.File]::ReadAllBytes($ccl.FullName)
$cclStr = [System.Text.Encoding]::ASCII.GetString($cclBin)
$expected_strings = @("NEEDS_VOL", "STEP-only", "radia_*.py")
foreach ($s in $expected_strings) {
  if ($cclStr -notmatch [regex]::Escape($s)) {
    $fail += "CCL MISSING STRING: '$s' not found in $($ccl.FullName)"
  }
}

# --- Check 3: import radia resolves to W:\..\src\radia (editable) ---
# editable install → samefile against the W: source.  If samefile is
# False, the editable was overwritten by a non-editable `pip install .`
# (the 2026-04-21 PEEC-inductance class of regression).
$check3 = python -c "import os, radia; print('OK' if os.path.samefile(radia.__file__, r'W:\00_CAE\Radia\01_GitHub\src\radia\__init__.py') else 'FAIL: ' + radia.__file__)" 2>$null
if ($check3 -notmatch "^OK") { $fail += "WRONG RADIA: $check3" }
$check3b = python -c "import os, cubit_mesh_export; print('OK' if os.path.samefile(cubit_mesh_export.__file__, r'W:\00_CAE\Radia\01_GitHub\packages\cubit-mesh-export\src\cubit_mesh_export\__init__.py') else 'FAIL: ' + cubit_mesh_export.__file__)" 2>$null
if ($check3b -notmatch "^OK") { $fail += "WRONG CME: $check3b" }

# --- Check 4: every radia_*.py present in W:\..\src\radia ---
# editable install means W: IS the install location, so just confirm
# the source files exist (no separate site-packages copy to enumerate).
$src_on_nas = "W:\00_CAE\Radia\01_GitHub\src\radia"
$missing = @()
foreach ($f in Get-ChildItem "$src_on_nas\radia_*.py" -File -EA SilentlyContinue) {
  if (-not (Test-Path -LiteralPath $f.FullName)) {
    $missing += $f.Name
  }
}
if ($missing) { $fail += "MISSING FROM W:\src\radia: $($missing -join ', ')" }

# --- Check 5: clear .pyc cache so next import re-compiles ---
# Editable install drops __pycache__ next to the .py files on W:.
$pycache = Join-Path $src_on_nas "__pycache__"
if (Test-Path $pycache) {
  Remove-Item -Recurse -Force $pycache -EA SilentlyContinue
  Write-Host "  cleared $pycache (forces fresh bytecode on next import)"
}
# Also clear panels/__pycache__ (editable picks up .py from W:\..\panels too)
$pycache2 = Join-Path $src_on_nas "panels\__pycache__"
if (Test-Path $pycache2) {
  Remove-Item -Recurse -Force $pycache2 -EA SilentlyContinue
  Write-Host "  cleared $pycache2"
}

# --- Check 6: NEEDS_VOL on new modules is actually enforced ---
# For each radia_*.py with NEEDS_VOL = False, confirm Python's module
# exposes it -- catches "file on disk but module stale in memory".
$peec_ok = python -c "import radia.radia_peec_inductance as m; assert m.NEEDS_VOL is False; print('OK')" 2>&1
if ($peec_ok -notmatch "OK") {
  $fail += "PEEC-inductance module did not expose NEEDS_VOL=False: $peec_ok"
}

# --- Check 7: Cubit plugin is still SymbolicLink (cubit-plugin-install
#     was NOT accidentally run, which would replace symlinks with
#     regular-file copies and break editable-deploy of the C++ side) ---
foreach ($p in @(
    "C:\Program Files\Coreform Cubit 2025.3\bin\radia_cubit.ccl",
    "C:\Program Files\Coreform Cubit 2025.3\bin\plugins\radia_cubit.ccm",
    "C:\Program Files\Coreform Cubit 2025.3\bin\plugins\radia_cubit_mesh.cp312-win_amd64.pyd")) {
  $f = Get-Item $p -EA SilentlyContinue
  if (-not $f) { $fail += "PLUGIN MISSING: $p"; continue }
  if ($f.LinkType -ne "SymbolicLink") { $fail += "REGULAR FILE (must be SymbolicLink): $p" }
}

# --- Report ---
if ($fail.Count -gt 0) {
  Write-Host ""
  Write-Host "=== DEPLOY VERIFICATION FAILED ==="
  $fail | ForEach-Object { Write-Host "  $_" }
  Write-Host ""
  Write-Host "DEPLOY IS NOT IN EFFECT.  Do NOT consider this deploy shipped."
  exit 1
}
Write-Host "OK: post-deploy verification passed -- deploy is in effect."
PS

If Check 1 fails, force-kill the zombie (step (a) was insufficient; rerun it). If Check 2 fails, the .ccl on W: is stale — rerun Build.ps1 on LAB; no copy/redeploy step is needed because the file on <Cubit>\bin\ is a symlink to W:. Check 3 catches a shadow Python install on PATH or an editable that was silently overwritten by a non-editable pip install . — repair via pip uninstall -y radia cubit-mesh-export followed by the first-time editable install in Step (b). Check 4-6 catch missing source files / stale bytecode / module-state-out-of-sync. Check 7 catches an accidental cubit-plugin-install run that replaced symlinks with regular-file copies — repair by re-creating the 3 symlinks per Step (b).

Do NOT declare the deploy complete until this step exits 0.

End-to-end functional test on 100号機

POLICY (2026-04-21): Every radia_*.py panel mode MUST run its canonical sample end-to-end AND the output numeric value must pass a golden-range check before the deploy is considered shipped. A subprocess that exits 0 is NOT enough — silently-wrong results (e.g. L off by 100x from a fallback path that accepted a bogus cross-section area) pass the status: "ok" check but the user sees garbage.

This closes the 2026-04-21 Kubota regression where my deploy pre-check all returned green (imports, hashes, QA, .ccl strings, NEEDS_VOL) but calc_peec_inductance had never been exercised on any real input file. The user ran the mode and got an unhandleable native crash on his first STEP.

Panel mode matrix

radia_*.py	calc_*.py	sample input	golden output range
radia_ih (BEM)	calc_heating_bem.py	ih_bem_sample.jou → .vol	P_total_W ≈ 11.55 ± 1
radia_ih (FEM-SIBC)	calc_fem_kelvin.py	ih_fem_kelvin_sample.jou → .vol	L ≈ 84 nH ± 5 (Cu 50 kHz)
radia_peec_inductance	calc_peec_inductance.py	ih_peec_inductance_coil.step (simple torus)	L ≈ 75 – 90 nH (Cu 50 kHz, n_peri=16)
radia_peec_inductance (multi-turn)	calc_peec_inductance.py	3turnCoil.jou (explicit centerline)	L ≈ 400 – 500 nH (Cu 150 kHz, n_peri=16)
radia_em	calc_em.py	em_sample.jou → .vol	(add range here when first run)
radia_pcb	calc_pcb.py	pcb_sample.jou → .vol	(add range here when first run)

When a new Simulink application mode is added, the deploy skill MUST be updated with a new matrix row BEFORE declaring the feature shipped. simulink-app-health and the relevant validation_test/ golden lane enforce this. New notebook workbenches are forbidden; IH is the temporary exception.

Every row must run on 100号機 (Stage 2) AND on mdx (Stage 3). Stage 1 (LAB) runs at least the rows for modes edited in this diff.

Sample E2E runner (scripted)

cat << 'PS' | ssh 100 'pwsh -Command -'
# Environment — editable install: $sp resolves to W:\..\src\radia
$sp = python -c "import os, radia; print(os.path.dirname(radia.__file__))"
$cubit = "C:\Program Files\Coreform Cubit 2025.3\bin\coreform_cubit.exe"
$tmp = "C:\temp"
if (-not (Test-Path $tmp)) { New-Item -Path $tmp -ItemType Directory | Out-Null }
$fails = @()

function Assert-InRange($value, $lo, $hi, $label) {
  if ($value -ge $lo -and $value -le $hi) {
    Write-Host ("  OK     {0}: {1:F3} (range [{2}, {3}])" -f $label, $value, $lo, $hi)
  } else {
    Write-Host ("  FAIL   {0}: {1:F3} NOT in [{2}, {3}]" -f $label, $value, $lo, $hi) -ForegroundColor Red
    $script:fails += "${label}: ${value} outside [${lo}, ${hi}]"
  }
}

# --- Row 1: PEEC-inductance, simple torus ---
Write-Host "`n=== PEEC-inductance: single-loop torus ==="
$step = Join-Path $sp 'panels\samples\ih_peec_inductance_coil.step'
if (-not (Test-Path $step)) {
  # Regenerate from .jou if the .step was not shipped (user regenerates on first play).
  $jou = Join-Path $sp 'panels\samples\ih_peec_inductance.jou'
  Copy-Item $jou $tmp\ih_peec_inductance.jou
  Push-Location $tmp
  & $cubit -batch -nographics -nojournal ih_peec_inductance.jou | Out-Null
  Pop-Location
  $step = Join-Path $tmp 'ih_peec_inductance_coil.step'
}
$out = python (Join-Path $sp 'panels\calc_peec_inductance.py') `
  --peec-step $step --peec-n-peri 16 --frequency 50000 `
  --current 1.0 --coil-sigma 5.8e7 2>&1
$json = ($out | Select-String '^\{' | Select-Object -Last 1).ToString()
$L_nH = ($json | ConvertFrom-Json).L_coil_nH
Assert-InRange $L_nH 75 90 "PEEC torus L_nH (50 kHz Cu)"

# --- Row 2: PEEC-inductance, multi-turn explicit-centerline .jou ---
Write-Host "`n=== PEEC-inductance: 3-turn explicit centerline ==="
$jou = Join-Path $sp 'panels\samples\3turnCoil.jou'
$out = python (Join-Path $sp 'panels\calc_peec_inductance.py') `
  --peec-step $jou --peec-n-peri 16 --frequency 150000 `
  --current 1.0 --coil-sigma 5.8e7 2>&1
$json = ($out | Select-String '^\{' | Select-Object -Last 1).ToString()
$L_nH = ($json | ConvertFrom-Json).L_coil_nH
Assert-InRange $L_nH 400 500 "PEEC 3-turn L_nH (150 kHz Cu)"

# --- Row 3: IH BEM sample (legacy) ---
Write-Host "`n=== IH BEM: gapped-torus sample ==="
$jou = Join-Path $sp 'panels\samples\ih_bem_sample.jou'
$wrapper = Join-Path $tmp 'test_ih.jou'
"playback `"$jou`"`r`nexport netgen `"$tmp/ih_test.vol`" order 1 overwrite`r`n" |
    Set-Content -Path $wrapper -Encoding ASCII
& $cubit -batch -nographics -nojournal $wrapper | Out-Null
$out = python (Join-Path $sp 'panels\calc_heating_bem.py') `
    --vol $tmp\ih_test.vol --wp-label workpiece `
    --coil-radius 0.110 --coil-current 100 --gap-deg 5 --frequency 50000 `
    --material steel --half-thickness 0.050 --esim-geometry cylinder `
    --impedance-model linear --mu-r 100 2>&1
$json = ($out | Select-String '^\{' | Select-Object -Last 1).ToString()
$P = ($json | ConvertFrom-Json).P_total_W
Assert-InRange $P 10 13 "IH BEM P_total_W"

# --- Row N: add new panel modes here ---

# --- Summary ---
if ($fails.Count -gt 0) {
  Write-Host "`n=== E2E PANEL MATRIX FAILED ===" -ForegroundColor Red
  $fails | ForEach-Object { Write-Host "  $_" -ForegroundColor Red }
  Write-Host "`nDEPLOY IS NOT IN EFFECT.  Panel output values outside golden range."
  exit 1
}
Write-Host "`nOK: all panel-mode E2E rows passed golden-range checks."
PS

Golden-range principle (2026-04-21):

• status: "ok" from a subprocess is NOT sufficient evidence the feature works. L = 4.78 nH vs L = 426 nH both exit 0.
• Every mode asserts the numeric OUTPUT falls in a range that distinguishes correct physics from a silently-wrong fallback path.
• Ranges are WIDE (±10-20 %) to tolerate mesh / algorithm noise but tight enough that order-of-magnitude regressions (like the longest-edge section() bug) FAIL loudly.
• Updating a range requires a line in the commit message explaining what physics / algorithm change justifies the new range.

Release (CI 通過後):

git push origin main
gh run list --limit 3   # wait for CI success

# cubit-mesh-export (if changed)
git tag cubit-mesh-export-vX.Y.Z
git push origin cubit-mesh-export-vX.Y.Z

# radia (if changed)
git tag vX.Y.Z
git push origin vX.Y.Z

# radia-mcp (if changed)
git tag radia-mcp-vX.Y.Z
git push origin radia-mcp-vX.Y.Z

Stage 3: 100号機 (PyPI install, multi-user lab seats) {#stage-3-100go-pypi}

POLICY FLIP 2026-05-01: This stage WAS "mdx PyPI" until 2026-05-01. The roles were swapped: 100号機 now hosts the PyPI installation (pip install radia==X.Y.Z) so that kubota and other lab members exercise the public distribution channel as their daily workflow, while mdx is the editable iteration sandbox (Stage 2). Cubit plugin on 100号機 now goes through cubit-plugin-install (regular-file deploy from PyPI wheel), NOT the symlink chain.

# One-time on each release: install latest published version on 100号機
ssh 100 'C:/Progra~1/Python312/python.exe -m pip install --upgrade radia cubit-mesh-export radia-mcp'
ssh 100 'cubit-plugin-install'

The full L0–L4 ladder (see Check Matrix) runs against this PyPI install. Any failure here BLOCKS the release for the lab — kubota / keiko / yano all run this pip install. The legacy mdx-PyPI instructions below are kept as a historical reference only (mdx is now editable, see Stage 2).

Legacy mdx PyPI instructions (HISTORICAL — pre-2026-05-01)

# (No longer used — mdx switched to editable.)
ssh mdx 'pip install --upgrade radia cubit-mesh-export radia-mcp'

# Deploy Cubit plugin + panels
ssh mdx 'cubit-plugin-install'

# Verify
cat << 'PY' | ssh mdx 'python -'
import os, radia
from cubit_mesh_export import __version__ as cm_ver
print(f"radia v{radia.__version__}")
print(f"cubit-mesh-export v{cm_ver}")
for name, path in [("ccm", r"C:\Program Files\Coreform Cubit 2025.3\bin\plugins\radia_cubit.ccm"),
                   ("ccl", r"C:\Program Files\Coreform Cubit 2025.3\bin\radia_cubit.ccl")]:
    sz = os.path.getsize(path) if os.path.isfile(path) else 0
    print(f"  {name}: {sz:,} bytes")
PY

Pre-PyPI deploy (wheel pipe, PyPI 公開前にテストしたい場合):

# cubit-mesh-export
cat packages/cubit-mesh-export/dist/cubit_mesh_export-*.whl | ssh mdx 'python -c "import sys; open(r\"C:\temp\cm.whl\",\"wb\").write(sys.stdin.buffer.read())"'
ssh mdx 'pip install --force-reinstall C:\temp\cm.whl'

# radia
cat dist/radia-*.whl | ssh mdx 'python -c "import sys; open(r\"C:\temp\radia.whl\",\"wb\").write(sys.stdin.buffer.read())"'
ssh mdx 'pip install --force-reinstall --no-deps C:\temp\radia.whl'

# Deploy plugin + panels
ssh mdx 'cubit-plugin-install'

---

Wheel Sanity Check (run before any PyPI release)

pip wheel / Build_Wheel.ps1 の出力を検証する。 orphan ファイル混入を防ぐ。

radia wheel

cd S:/Radia/01_GitHub
rm -rf src/radia.egg-info
pip wheel . --no-deps --no-build-isolation --no-cache-dir -w C:/temp/wheel_check

python -c "
import zipfile, glob, sys
whl = glob.glob('C:/temp/wheel_check/*.whl')[0]
z = zipfile.ZipFile(whl)
names = z.namelist()
print(f'Wheel: {whl.split(\"/\")[-1]} ({len(names)} files)')

# 1. Must contain
must = ['radia/_radia_pybind.pyd', 'radia/radia_cubit.ccm', 'radia/radia_cubit.ccl',
        'radia/panels/panel_registry.json', 'radia/panels/calc_mesh_eval.py',
        'radia/panels/samples/ih_bem_sample.jou', 'radia/panels/samples/ih_bem_sample_B.sol',
        'radia/gmsh_post_export.py', 'radia/radia_gui_base.py']
ok = True
for m in must:
    found = any(m in n for n in names)
    if not found: print(f'  MISSING: {m}'); ok = False
if ok: print('  All required files present')

# 2. Must NOT contain (orphans)
bad = [n for n in names if ('/_' in n and n != 'radia/__init__.py'