Windows is supported: promote the job to required and say so

The experimental job went green at 177/177, so it is no longer
experimental. README, CHANGELOG, ROADMAP and docs/installation.md now say
Windows is CI-verified, with the platform notes that matter: .dll naming
without the lib prefix, the DLL-in-use lock that the unique-copy load
policy already sidesteps, and the LF requirement.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DPAgXWnvgpKcUw71GHStXC

Author: pegeaether

.github/workflows/ci.yml

@@ -66,9 +66,8 @@ jobs:
       - run: make test-m4
 
   sbcl-windows:
-    name: SBCL / Windows x86-64 (experimental, may fail)
+    name: SBCL / Windows x86-64 (required)
     runs-on: windows-latest
-    continue-on-error: true
     defaults:
       run:
         shell: bash
@@ -83,9 +82,6 @@ jobs:
           curl -sO https://beta.quicklisp.org/quicklisp.lisp
           sbcl --non-interactive --load quicklisp.lisp \
                --eval '(quicklisp-quickstart:install)'
-      # The loader was ported (LoadLibrary/GetProcAddress, pointer-sized
-      # uintptr for LLP64, .dll naming with no lib prefix) but has never run
-      # on Windows — this job is where that claim gets tested.
       - name: full gate
         run: make test-m4
 

CHANGELOG.md

@@ -17,6 +17,10 @@ system. The C ABI has its own version, checked at load time: **ABI 1 since
   `with-client`), a hermetic loopback test server, and `audit.sh`, an
   executable BOUNDARY §7 check.
 - `make bench` — a benchmark suite for the boundary paths.
+- **Windows support.** The loader is abstracted over `dlopen`/`dlsym` and
+  `LoadLibrary`/`GetProcAddress`, `uintptr` is derived from the pointer
+  size (naming a C type is wrong on LLP64), and artifact naming knows
+  Windows drops the `lib` prefix. CI runs the full suite there, 177/177.
 
 ### Changed
 - **Bulk `:bytes`/`:vec` marshalling**: pinned vector + `memcpy`, with the
@@ -28,6 +32,10 @@ system. The C ABI has its own version, checked at load time: **ABI 1 since
   rejected before anything is interned.
 - `#[rulisp(constructor)]` on a `&self` method dropped the receiver and
   surfaced as a raw `E0061`; now a compile error naming the fix.
+- The `Cargo.toml` scraper could not tolerate CR, so `use-crate` failed on
+  any CRLF manifest. The tree is also pinned to LF now: a CRLF checkout
+  broke the golden byte-identity fixture and turned Lisp format-string
+  line continuations into `FORMAT-ERROR`.
 
 ## 0.2.1 — 2026-08-03
 

ROADMAP.md

@@ -48,14 +48,20 @@ providing conditions, restarts and `with-client`. Headline finding:
   audit-gate regex that could never fire because glibc imports are
   version-tagged, and a `TaskGuard` ordering that published "done" before
   releasing the admission permit.
-- ✅ **Windows groundwork**: `uintptr` is now derived from the pointer size
-  (naming a C type was wrong on LLP64, where `unsigned long` is 32 bits);
-  the loader is abstracted over `dlopen`/`dlsym` and
-  `LoadLibrary`/`GetProcAddress`; artifact naming knows that Windows drops
-  the `lib` prefix and uses `.dll`. The unique-copy load policy already
-  sidesteps the DLL-in-use lock. **None of this has run on Windows** — an
-  experimental (`continue-on-error`) CI job is where the claim gets tested.
-  Finishing the port is a v0.4 goal.
+- ✅ **Windows works** — and earlier than planned. `uintptr` is derived
+  from the pointer size (naming a C type was wrong on LLP64, where
+  `unsigned long` is 32 bits); the loader is abstracted over
+  `dlopen`/`dlsym` and `LoadLibrary`/`GetProcAddress`; artifact naming
+  knows Windows drops the `lib` prefix and uses `.dll`; the unique-copy
+  load policy, added for macOS dyld caching, also sidesteps the DLL-in-use
+  lock. Three real portability bugs surfaced on the way, none of them
+  Windows-only in principle: a Cargo.toml scraper that could not tolerate
+  CR, the golden fixture broken by CRLF translation, and Lisp format
+  strings whose `~` end-of-line continuation is not the tilde-newline
+  directive once a CR sits between them (the tree is now pinned to LF).
+  The CI job is **required**, at 177/177. (It runs three fewer assertions
+  than Linux: `fx.target-check` guards a couple of them behind
+  `#+(and x86-64 linux)`.)
 - ✅ A worked Deploy recipe (docs/distribution.md), including the two things
   that bite silently: platform-named artifacts and quiescing foreign
   threads in a dump hook.

docs/installation.md

@@ -73,6 +73,24 @@ macOS notes:
 - rulisp loads libraries by absolute path, so SIP's stripping of
   `DYLD_LIBRARY_PATH` is irrelevant.
 
+## Windows
+
+> Status: CI-verified on Windows x86-64 with SBCL (full suite green).
+
+Install [SBCL](http://www.sbcl.org/platform-table.html) (or
+`choco install sbcl`), a Rust toolchain from <https://rustup.rs>, and
+Quicklisp as above. Notes:
+
+- Artifacts are `<name>.dll` — cargo drops the `lib` prefix on Windows —
+  and `load-blob-crate` looks for `<name>-windows-x86_64.dll`.
+- A loaded DLL is locked, so an artifact cannot be overwritten in place.
+  rulisp already loads a uniquely named copy per generation, so
+  `use-crate`/`reload-crate` work anyway; stale copies in the cache are
+  swept on a later run.
+- Check the repository out with LF endings (the shipped `.gitattributes`
+  does this): the manifest fixtures are byte-compared, and a CR inside a
+  Lisp format string breaks its line continuations.
+
 ## Getting rulisp
 
 Until the Quicklisp/Ultralisp registration lands, clone into Quicklisp's