diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 3d87a50..900b5c5 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -48,6 +48,46 @@ jobs:
       - name: Run tests
         run: cargo test --workspace --all-targets
 
+  # The same checks on the other two operating systems the CLI is meant to run
+  # on. Until this job existed no `cfg(not(unix))` branch had ever been
+  # compiled, let alone linted or run. See
+  # docs/plans/cross-platform-plan-2026-09-14.md, phase 0.
+  #
+  # A separate job rather than a matrix on `rust` above, for two reasons.
+  # Formatting cannot vary by platform, so it runs once. And branch protection
+  # requires a check named exactly "Rust": a matrix would rename it and block
+  # every pull request until the settings changed. Add these two to the
+  # required checks once they are green.
+  rust-platforms:
+    name: Rust (${{ matrix.os }})
+    runs-on: ${{ matrix.os }}
+    # Windows and macOS runners start slower and build slower than Linux.
+    # Set from what the first runs take, not from the Linux job's 15.
+    timeout-minutes: 40
+    strategy:
+      fail-fast: false
+      matrix:
+        os:
+          - macos-14
+          - windows-2025
+    steps:
+      - name: Check out source
+        uses: actions/checkout@v7
+        with:
+          persist-credentials: false
+      - name: Install Rust
+        uses: dtolnay/rust-toolchain@stable
+        with:
+          components: clippy
+      - name: Run Clippy
+        run: cargo clippy --workspace --all-targets --all-features -- -D warnings
+      # Runs even when Clippy fails. A lint and a failing test are separate
+      # findings, and one run should report both rather than hiding the tests
+      # behind the first warning.
+      - name: Run tests
+        if: success() || failure()
+        run: cargo test --workspace --all-targets
+
   web:
     name: Web
     runs-on: ubuntu-24.04
diff --git a/cli/tests/archive.rs b/cli/tests/archive.rs
index 17801b8..cad992c 100644
--- a/cli/tests/archive.rs
+++ b/cli/tests/archive.rs
@@ -11,9 +11,12 @@ use std::{
 };
 
 use drop_cli::{
-    tar::{TarPlan, safe_relative_path, symlink_target_stays_inside, traverses_only_real_dirs},
+    tar::{TarPlan, safe_relative_path, symlink_target_stays_inside},
     untar::TarExtractor,
 };
+// Only the symlink tests use this, and planting a symlink needs Unix.
+#[cfg(unix)]
+use drop_cli::tar::traverses_only_real_dirs;
 
 fn scratch(name: &str) -> PathBuf {
     let base = std::env::temp_dir().join(format!(
diff --git a/cli/tests/transfer.rs b/cli/tests/transfer.rs
index 7d355bb..b6cd155 100644
--- a/cli/tests/transfer.rs
+++ b/cli/tests/transfer.rs
@@ -318,6 +318,10 @@ async fn a_malformed_code_fails_before_the_relay_is_contacted() {
 }
 
 /// Builds a ustar archive from `(name, typeflag, link_target, contents)`.
+///
+/// Unix only because its one caller is: the hostile archive it builds plants
+/// symlinks, and the receiver cannot create those on Windows.
+#[cfg(unix)]
 fn archive_of(entries: &[(&str, u8, &str, &[u8])]) -> Vec<u8> {
     let mut archive = Vec::new();
 
diff --git a/docs/plans/cross-platform-plan-2026-09-14.md b/docs/plans/cross-platform-plan-2026-09-14.md
index cb318f7..33c7f39 100644
--- a/docs/plans/cross-platform-plan-2026-09-14.md
+++ b/docs/plans/cross-platform-plan-2026-09-14.md
@@ -1,6 +1,6 @@
 # Cross-platform plan: Windows, macOS and Linux, and transfers between them
 
-Status: **proposed**
+Status: **active** — phase 0 done 2026-09-14
 Created: **2026-09-14**
 Last updated: **2026-09-14**
 
@@ -181,20 +181,50 @@ where this plan spends its effort.
 Nothing below can be verified until this exists, so it comes first and its
 failures are the input to the rest.
 
-- [ ] `ci.yml`: turn the `rust` job into a matrix over `ubuntu-24.04`,
-      `macos-14` (arm64) and `windows-2025`. Formatting stays on Linux only,
-      because it cannot vary by platform. Clippy runs on all three, since
-      `cfg(windows)` code is otherwise never linted.
-- [ ] Record what fails on the first run in this plan, dated, before fixing
-      any of it.
-- [ ] Make the suite pass on all three without skipping anything that is not
-      genuinely inapplicable. Every new `cfg` on a test gets a comment saying
-      why the test cannot run there.
-- [ ] Mark `.github/workflows/ci.yml` checkouts `autocrlf`-safe:
-      `.gitattributes` already sets `eol=lf`, so fixture bytes should match
-      across platforms. Confirm on the Windows runner rather than assume.
-
-**Gate:** `cargo test --workspace --all-targets` green on all three runners.
+- [x] `ci.yml`: a `rust-platforms` job running Clippy and the tests on
+      `macos-14` and `windows-2025`. It is a separate job, not a matrix on
+      `rust`, because branch protection requires a check named exactly "Rust"
+      and a matrix would rename it. Formatting stays on Linux only. The test
+      step runs even when Clippy fails, so one run reports both.
+- [x] Record what fails on the first run in this plan, dated, before fixing
+      any of it. See below.
+- [x] Make the suite pass on all three without skipping anything that is not
+      genuinely inapplicable. Each new `cfg` has a comment saying why.
+- [x] Line endings: no fixture failed on Windows, so `.gitattributes`' `eol=lf`
+      holds for what the tests read. Confirmed by the Windows run, not assumed.
+- [ ] Add `Rust (macos-14)` and `Rust (windows-2025)` to branch protection's
+      required checks. A repository setting, for the owner.
+
+#### What the runners found, 2026-09-14
+
+**Run 1.** macOS: Clippy and all tests green on the first attempt. Windows:
+Clippy failed on two items, and nothing else. `cli/tests/archive.rs` imported
+`traverses_only_real_dirs`, and `cli/tests/transfer.rs` defined `archive_of`.
+Both are used only by `#[cfg(unix)]` symlink tests. **Every `cfg(not(unix))`
+branch in the CLI compiled and passed Clippy**, the first time any of it had
+been compiled. Fixed by gating the two items. The test step had not run,
+because it followed Clippy.
+
+**Run 2.** Windows: Clippy and **the whole suite green**, about 6 minutes on a
+cold runner. macOS: one failure,
+`upload_socket_rejects_chunks_over_the_message_cap`, with
+`expected oversized chunk to be written: Io(... BrokenPipe ...)`. It passed on
+run 1, so it is a race. The relay refuses an oversized frame from its header
+and closes the socket, possibly while the test is still writing the frame body.
+Linux's socket buffer usually absorbs the write, and macOS's usually does not.
+The relay's behaviour was correct both times; the test demanded that its own
+doomed write succeed. The test now accepts a broken pipe, reset or closed
+connection on that write and still asserts what matters: the receiver gets an
+error and never a byte of the chunk.
+
+**What green does not mean.** Windows passing the suite says the suite has
+nothing Windows-specific in it. It does not say Windows works. Findings 1–5
+above have no tests at all: the seven symlink tests are Unix-only, and nothing
+exercises reserved names, alternate data streams, trailing dots, or a symlink
+arriving at a Windows receiver. Phase 1 writes those tests, and they are
+expected to fail before its fixes.
+
+**Gate:** met, pending the final run with the race fixed.
 
 ### Phase 1 — a Windows receiver writes what it can and reports the rest
 
diff --git a/tests/websocket_transfer.rs b/tests/websocket_transfer.rs
index be4e733..6477dfc 100644
--- a/tests/websocket_transfer.rs
+++ b/tests/websocket_transfer.rs
@@ -162,10 +162,34 @@ async fn upload_socket_rejects_chunks_over_the_message_cap() {
         .expect("expected sender meta message to be sent");
     next_json_message_matching(&mut receiver_ws, |payload| payload["type"] == "meta").await;
 
-    sender_ws
-        .send(Message::binary(oversized))
-        .await
-        .expect("expected oversized chunk to be written");
+    // The relay refuses the frame from its header and closes the socket, which
+    // can happen while this side is still writing the frame's body. Whether
+    // the write finishes first depends on how much the kernel's socket buffer
+    // absorbs: on Linux it usually does, on macOS it usually does not, and the
+    // write fails with a broken pipe or a reset. Both mean the relay hung up on
+    // an oversized frame, which is the behaviour under test. What is asserted
+    // is what the receiver saw, below.
+    if let Err(error) = sender_ws.send(Message::binary(oversized)).await {
+        let hung_up = matches!(
+            &error,
+            tokio_tungstenite::tungstenite::Error::Io(io)
+                if matches!(
+                    io.kind(),
+                    std::io::ErrorKind::BrokenPipe
+                        | std::io::ErrorKind::ConnectionReset
+                        | std::io::ErrorKind::ConnectionAborted
+                )
+        ) || matches!(
+            &error,
+            tokio_tungstenite::tungstenite::Error::ConnectionClosed
+                | tokio_tungstenite::tungstenite::Error::AlreadyClosed
+        );
+
+        assert!(
+            hung_up,
+            "writing the oversized chunk failed for a reason other than the relay hanging up: {error}"
+        );
+    }
 
     // The relay must tear the session down instead of relaying the chunk. The
     // receiver stream therefore ends, with an error frame rather than any part