Refactor: descriptors, one Step runner, vocabulary and translation guards (#17, #18, #20, #21, #22)

[DireDoch]

Implements #17, #18, #20, #21 and #22, in the order their dependencies impose:
#22 → #21 → #20 → #17 → #18. One commit per issue.

#19 is not implemented — see the note at the bottom.

What each commit does

Issue	Commit	The point
#22	82d9d82	Test-WcdNetworkAdapterActive folded into its callers; the virtual-adapter exclusion list promoted to a named script-scope variable at the top of the file, where someone whose VPN client turned up in the summary will find it.
#21	91d5921	The Form Factor / Environment / Optional Tool drift renamed out of Resolve-WcdExecutionOptions and both $T tables; $moduleStatus loses the three French identifiers. The extended stale-vocabulary pattern is the actual deliverable.
#20	c5923f1	tests/Translations.Tests.ps1 asserts the EN and FR tables expose identical keys, recursing into Checklist, Labels, Remedy and ComputerNameRejected.
#17	5da3ed2	Every Config-*.ps1 exports a descriptor. The $modules array and the switch ($modName) dispatch are gone; the orchestrator globs src/Config-*.ps1.
#18	d7e1394	Invoke-WcdStep owns the progress events, both log lines, the try/catch and the Result shape. Config-Power and Config-TaskbarLeft converted; the other ten left alone deliberately.

Two issues disagreed with the code, and the code won

• Fold the single-caller network adapter predicates into their caller #22 says Test-WcdNetworkAdapterActive has one caller. It has two —
  Config-Network.ps1:113 and :519. The regex is inlined at both.
• Modules declare themselves with a descriptor instead of six registration edits #17 says the checklist is "in run order". It is not, and never was:
  Modules run Identity → Power → Decimal → Taskbar → …, the checklist reads
  Identity → Restart → Taskbar → Language → Keyboard → Decimal → Power → ….
  So a descriptor carries Order and RowOrder. Both live in the Module's
  own file, so it is still one registration point rather than six.

How "byte-identical" was verified

Not by reading the diff. Two harnesses dot-source both trees and compare:

• Modules declare themselves with a descriptor instead of six registration edits #17 — progress plan, every checklist row (Step, Label, Kind,
  Detail) and the rendered line, for 7 scenarios × 2 languages:
  laptop/desktop, workstation/vdi, declined apps, identity applied, update
  reboot pending, no printers, winget missing, nothing ran, unelevated power.
  554 lines, zero diff.
• Invoke-WcdStep: one Step runner behind every Module #18 — every Result property, every log line (timestamps stripped) and
  every progress event, for 8 paths: laptop/desktop × elevated/unelevated,
  powercfg throwing, registry write succeeding / denied / failing generically.
  58 lines, zero diff — including property order, which the JSON report
  reproduces and which needed [ordered] on the Power fragment to preserve.

Each new guard was also verified by breaking the thing it guards:

• Rename the CONTEXT.md terms that drifted inside the implementation #21 — reintroduced $deviceResult, the Etapes hashtable key and
  $ModuleStatus.Etapes one at a time; each fails the test.
• Test that the EN and FR translation tables expose identical keys #20 — deleted a top-level FR key, a nested FR key and a nested EN key; each
  fails, in both directions.
• Modules declare themselves with a descriptor instead of six registration edits #17 — Test-WcdModuleDescriptor is itself tested against each missing field.

Also in here

• CLAUDE.md: how to run pwsh and Pester on Linux without sudo, and which
  two tests fail there on purpose (New-WinUserLanguageList and
  Get-CimInstance cannot be mocked away — Pester's Mock needs the command to
  exist first). A clean local run is 165 passed / 2 failed; CI on Windows is
  still the authority.
• Manual chapter 7 rewritten — registration is one code block in one file — and
  a new section on when Invoke-WcdStep is the wrong tool. docs/manual.pdf
  recompiled.
• Dropped as dead: $T.ModuleNotFound in both tables (a globbed file cannot be
  missing) and the PrinterAdd / PrinterSkip step labels, which no Module
  emitted.
• .gitignore picks up docs/research/, which was already uncommitted in the
  working tree.

#19 is deliberately not here

Its own author writes that it "should be closed as wontfix without
embarrassment if it keeps losing to more useful work", and that it is worth
doing for tidiness rather than for extensibility. #20 is what actually stops the
bug, and it landed. Closing #19 as wontfix for the reason the issue already
gives.

🤖 Generated with Claude Code