Add native Windows support: install.ps1, doctor.ps1, and Windows settings

.github/workflows/ci.yml

@@ -12,11 +12,30 @@ jobs:
       - name: Shell scripts parse
         run: bash -n install.sh tools/doctor.sh bench/run.sh bench/task/verify.sh
 
+      - name: PowerShell scripts parse
+        shell: pwsh
+        run: |
+          $bad = @()
+          foreach ($f in 'install.ps1', 'tools/doctor.ps1') {
+            $tokens = $null; $errors = $null
+            [System.Management.Automation.Language.Parser]::ParseFile((Resolve-Path $f), [ref]$tokens, [ref]$errors) | Out-Null
+            if ($errors) { $bad += "${f}:"; $bad += $errors }
+          }
+          if ($bad) { $bad | Write-Host; exit 1 }
+          Write-Host 'powershell scripts parse'
+
       - name: Python components compile
         run: python3 -m py_compile claude/hooks/*.py claude/cli/*.py bench/score.py bench/task/loglib/*.py
 
-      - name: Settings snippet is valid JSON
-        run: python3 -c "import json; json.load(open('claude/settings/settings-snippet.json'))"
+      - name: Settings snippets are valid JSON (all four)
+        run: |
+          python3 - <<'EOF'
+          import json
+          for name in ("settings-snippet.json", "settings-snippet-small.json",
+                       "settings-snippet-windows.json", "settings-snippet-windows-small.json"):
+              json.load(open(f"claude/settings/{name}"))
+          print("snippets ok")
+          EOF
 
       - name: Workflow scripts compile
         run: node tools/check-workflows.mjs
@@ -58,3 +77,45 @@ jobs:
           total=$(/tmp/bench-venv/bin/python bench/score.py /tmp/pristine-inst /tmp/bench-venv/bin/python \
             | python3 -c "import json,sys; print(json.load(sys.stdin)['total'])")
           test "$total" = "1" || { echo "pristine anchor drifted: scored $total/15, expected 1/15"; exit 1; }
+
+  windows:
+    runs-on: windows-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Python components compile
+        shell: pwsh
+        run: |
+          $files = Get-ChildItem claude/hooks/*.py, claude/cli/*.py | ForEach-Object FullName
+          python -m py_compile @files
+          if ($LASTEXITCODE -ne 0) { exit 1 }
+
+      - name: Install script end-to-end (twice, idempotent)
+        shell: pwsh
+        run: |
+          $env:CLAUDE_DIR = Join-Path $env:RUNNER_TEMP 'fake-claude'
+          ./install.ps1 *>&1 | Out-Null
+          # install.ps1 reports via Write-Host (information stream) — merge all
+          # streams or the capture is empty and this check can never pass.
+          $second = ./install.ps1 *>&1 | Out-String
+          if ($second -notmatch 'unchanged') { Write-Host 'second run not idempotent'; exit 1 }
+          foreach ($rel in 'skills/fable/SKILL.md', 'skills/webdesign/references/german-market.md', 'cli/mem.py') {
+            if (-not (Test-Path (Join-Path $env:CLAUDE_DIR $rel))) { Write-Host "missing: $rel"; exit 1 }
+          }
+          # backups must never create loadable duplicates inside skills/
+          $installed = (Get-ChildItem (Join-Path $env:CLAUDE_DIR 'skills')).Count
+          $shipped = (Get-ChildItem claude/skills).Count
+          if ($installed -ne $shipped) { Write-Host "skills dir has $installed entries, expected $shipped"; exit 1 }
+
+      - name: Doctor verifies the merged install
+        shell: pwsh
+        run: |
+          $env:CLAUDE_DIR = Join-Path $env:RUNNER_TEMP 'fake-claude'
+          Copy-Item claude/settings/settings-snippet-windows.json (Join-Path $env:CLAUDE_DIR 'settings.json')
+          ./tools/doctor.ps1
+
+      - name: Unit tests (POSIX-installer tests self-skip)
+        shell: pwsh
+        run: |
+          python -m pip -q install pytest
+          python -m pytest tests/ -q

CHANGELOG.md

@@ -1,5 +1,56 @@
 # Changelog
 
+## v1.9 — 2026-07-09
+
+Windows pass + README redesign. The kit's discipline layer was always OS-portable — the
+hooks and the mem CLI are stdlib Python with `expanduser`/`os.path` throughout, and the
+privacy guard already probed for case-insensitive filesystems — but the delivery layer
+(bash installer, bash doctor, `python3 ~/...` hook commands) was POSIX-only, so on native
+Windows the kit was silently uninstallable. This release makes Windows a first-class
+install target with the same no-silently-inert guarantees, and restructures the README
+around the reader (install first, story second, inventory collapsible).
+
+### Added
+- **`install.ps1` — native Windows installer, full parity with `install.sh`.** Same
+  out-of-tree backups, same hash-based idempotency, same `.fable-manifest` skill
+  tracking with stale-file pruning, same `-Tier small` / `-StrongModel <m>` flags,
+  same never-edit-settings posture. Parity is enforced, not asserted: tests pin the
+  skill manifests byte-for-byte across both installers and require that running
+  `install.ps1` over a bash-installed tree reports everything unchanged (a dual-boot /
+  WSL+native machine must never churn backups). Python launcher discovery tries
+  `py -3`, `python`, `python3` in order and soft-fails like the bash bootstrap.
+- **`tools/doctor.ps1` — native Windows doctor, same checks and exit codes as
+  `doctor.sh`**, plus two Windows-only diagnoses: Git Bash present (on native Windows
+  it is the hook command shell — absent means every hook is inert, a FAIL), and a
+  warning when `settings.json` wires hooks through `python3` (a Unix snippet merged on
+  Windows never fires — the exact silently-inert failure the doctor exists to catch).
+- **Windows settings snippets** (`settings-snippet-windows.json`,
+  `-windows-small.json`): identical to the Unix snippets except hook commands invoke
+  `python` (Windows Pythons ship no `python3` launcher). `tests/test_windows_port.py`
+  keeps them in lockstep structurally — any drift from the tested Unix configuration
+  fails CI.
+- **`tests/test_windows_port.py`** — snippet-mirror guards (run everywhere) plus
+  pwsh-gated end-to-end tests mirroring `test_install_doctor.py`: install twice
+  (idempotent, no backup churn), user files in skill dirs preserved, formerly-shipped
+  files pruned, strong-model pin byte-identical with bash, doctor pass/fail/unwired
+  scenarios. The POSIX-installer tests now self-skip on Windows instead of driving
+  bash scripts through Git Bash.
+- **CI `windows` job** (`windows-latest`): compiles every Python component, runs
+  `install.ps1` end-to-end twice, verifies the merged install with `doctor.ps1`, and
+  runs the unit suite. The Ubuntu job additionally parse-checks both `.ps1` scripts
+  and validates all four settings snippets.
+
+### Changed
+- **README redesigned.** Install (macOS/Linux and Windows side by side, with a
+  collapsible Windows-notes block) now leads; the origin story is two paragraphs, not
+  a wall; the full annotated file tree is collapsible behind a six-row component-layer
+  table; badges + section nav on top. Every honest-limits paragraph survives, plus a
+  new one: the Windows port is CI-verified end-to-end but has not yet had a live
+  Claude Code session pass on native Windows.
+- Shipped components that name the mem CLI (`claude/CLAUDE.md` doctrine line,
+  `memory-search` skill, `/memory-gc` agent prompts) now note the Windows spelling:
+  `python` wherever a command says `python3`.
+
 ## v1.8 — 2026-07-08
 
 Memory pass: the kit stops forgetting across projects. Native auto-memory is per-git-repo,