diff --git a/.github/workflows/check.yaml b/.github/workflows/check.yaml
index ba13f09..2d72fc4 100644
--- a/.github/workflows/check.yaml
+++ b/.github/workflows/check.yaml
@@ -23,8 +23,28 @@ env:
 
 jobs:
   check:
-    runs-on: ubuntu-latest
-    timeout-minutes: 25
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: ${{ matrix.timeout }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - os: ubuntu-latest
+            timeout: 25
+          - os: macos-latest
+            timeout: 45
+          - os: windows-latest
+            timeout: 75
+    # Windows runs the checkers under busybox-w32 ash, same as test.yaml's Windows lane.
+    # That's the shell mise's `shell = "bash"` tasks run on; Git Bash's MSYS fork-emulation cygheap breaks
+    # down on long task chains. The path matches the version pin in `.mise/config.windows.toml`'s
+    # [tools."http:busybox"] entry.
+    defaults:
+      run:
+        shell: >-
+          ${{ startsWith(matrix.os, 'windows') &&
+              'C:\Users\runneradmin\AppData\Local\mise\installs\http-busybox\1.37.0\ash.exe -euo pipefail {0}'
+              || 'bash --noprofile --norc -euo pipefail {0}' }}
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -32,8 +52,13 @@ jobs:
           fetch-depth: 1
           persist-credentials: false
 
+      - name: Free disk space on Windows
+        if: runner.os == 'Windows'
+        uses: ./.github/actions/free-disk-space-windows
+
       - name: Install mise + tools
         uses: ./.github/actions/install-mise-tools
+        timeout-minutes: 25
         with:
           github-token: ${{ github.token }}
 
@@ -43,6 +68,15 @@ jobs:
           GITHUB_TOKEN: ${{ github.token }}
           CARGO_NET_RETRY: "5"
 
+      # MISE_JOBS=1 serializes mise's task scheduler so the parallel cargo checkers don't race on Windows.
+      # cargo-check / cargo-clippy-check / cargo-doc-check otherwise contend on cargo's
+      # `.cargo/registry/index` lock -- on Linux/macOS cargo waits gracefully, on Windows it fails fast
+      # (see test.yaml's matching step). mise rejects an empty MISE_JOBS, and GHA has no `if:` on
+      # individual env entries, so it's a separate conditional step.
+      - name: Serialize mise tasks on Windows
+        if: runner.os == 'Windows'
+        run: echo "MISE_JOBS=1" >> "$GITHUB_ENV"
+
       - name: Run checkers
         run: |
           mise run check
diff --git a/.mise/config.dart.toml b/.mise/config.dart.toml
index 52e977e..9cd35e8 100644
--- a/.mise/config.dart.toml
+++ b/.mise/config.dart.toml
@@ -98,11 +98,7 @@ shell = "bash -euo pipefail -c"
 [tasks."gen:dart-ws"]
 depends = ["gen:ws-spec"]
 description = "Emit the Dart sealed-class WS client via dart-typegen (consumes generated/specs/ws.kdl from gen:ws-spec)"
-# Skip the regen on Windows; the committed file stays canonical.
-# That committed file is regenerated on Linux/macOS. dart-typegen exits 0 on Windows but writes an empty .dart
-# file, which would clobber the committed source and fail gen-specs-check.
 run = """
-[ "${OS:-}" = "Windows_NT" ] && { echo "gen:dart-ws: skipped on Windows (dart-typegen empty output)"; exit 0; }
 mkdir -p generated/dart-ws/lib
 dart-typegen generate -i generated/specs/ws.kdl -o generated/dart-ws/lib/ws_messages.dart
 dart format generated/dart-ws/lib/ws_messages.dart
@@ -113,10 +109,10 @@ shell = "bash -euo pipefail -c"
 depends = ["gen:ws-spec"]
 description = "Emit the Dart REST client via swagger_parser + build_runner (consumes generated/specs/rest.yaml)"
 dir = "generated/dart-rest"
-# Regenerate the Dart REST client, handling a Windows skip plus two swagger_parser output fixups.
-# Skipped on Windows like gen:dart-ws -- build_runner/dart-format output drifts there and gen-specs-check is
-# Windows-skipped anyway, so the Linux/macOS-regenerated files stay canonical. `rm -rf lib` first so a renamed
-# output (e.g. client_postfix change) can't leave orphaned files behind.
+# Regenerate the Dart REST client, applying two swagger_parser output fixups.
+# `rm -rf lib` first so a renamed output (e.g. client_postfix change) can't leave orphaned files behind. The
+# goawk rewrites pipe through `tr -d '\r'` because goawk emits CRLF on Windows and dart format preserves
+# whatever line endings it finds.
 #
 # Fixup 1 (storage.dart): swagger_parser renders the application/octet-stream storage endpoints unusably for the
 # browser -- a `dart:io` `File` PUT body and a `void` GET response. goawk rewrites it before build_runner so putFile
@@ -127,7 +123,6 @@ dir = "generated/dart-rest"
 # root client's sub-client imports as `clients/<name>_.dart` -- a stray trailing underscore that export.dart does
 # not have. goawk strips it. Idempotent: a no-op if upstream fixes the inconsistency.
 run = """
-[ "${OS:-}" = "Windows_NT" ] && { echo "gen:dart-rest: skipped on Windows"; exit 0; }
 dart pub get
 coreutils rm -rf lib
 dart run swagger_parser
@@ -148,9 +143,10 @@ END {
     printf("gen:dart-rest storage fixup miss: io=%d get=%d body=%d\\n", io, get, body) > "/dev/stderr"
     exit 1
   }
-}' lib/clients/storage.dart > "$s/storage.dart"
+}' lib/clients/storage.dart | coreutils tr -d '\\r' > "$s/storage.dart"
 coreutils mv "$s/storage.dart" lib/clients/storage.dart
-goawk '/^import / && /clients/ { gsub(/_\\.dart/, ".dart") } { print }' lib/rest_client.dart > "$s/rest_client.dart"
+goawk '/^import / && /clients/ { gsub(/_\\.dart/, ".dart") } { print }' lib/rest_client.dart |
+  coreutils tr -d '\\r' > "$s/rest_client.dart"
 coreutils mv "$s/rest_client.dart" lib/rest_client.dart
 dart run build_runner build
 dart format lib
diff --git a/.mise/config.toml b/.mise/config.toml
index b5b82f7..830ffc3 100644
--- a/.mise/config.toml
+++ b/.mise/config.toml
@@ -1040,15 +1040,7 @@ description = "Fail if any checked-in artifact under generated/ is stale"
 # All et-int-gen outputs land under generated/.
 # The runner's wit/deps/et-ws-messages is a symlink into that tree, so a single diff
 # against generated/ covers everything.
-#
-# Skipped on Windows: the regen pipeline produces non-deterministic drift there (nightly rustfmt formats
-# long signatures and doc-comment indentation differently across nightlies; pipx-driven python codegen
-# emits subtly different output). Linux/macOS runs are canonical.
-run = """
-[ "${OS:-}" = "Windows_NT" ] && { echo "gen-specs-check: skipped on Windows (non-deterministic)"; exit 0; }
-git diff --exit-code -- generated
-"""
-shell = "bash -euo pipefail -c"
+run = "git diff --exit-code -- generated"
 
 # `cargo-watch` is unmaintained; `watchexec` is its successor.
 # watchexec is what `cargo-watch` always called under the hood. Restart on .rs/.toml changes so a
diff --git a/CLAUDE.md b/CLAUDE.md
index a461a6b..03b014d 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -430,22 +430,6 @@ merged). If this signature recurs, stop rerunning and fix the root cause: raise
 runner-registration timeout in the pyo3-runner module tests, or warm the torch import before the registration clock
 starts.
 
-### Known intermittent CI failure: wasi-runner vector OTLP store-and-forward timeout
-
-`et-ws-wasi-runner`'s `vector_otlp_relay::vector_relays_buffered_otlp_after_backend_comes_online` intermittently
-fails on the Windows `override` job (test.yaml) with the literal failure line
-
-    mock never received the relayed `relay-probe` span -- store-and-forward failed
-
-from `services/ws-wasi-runner/tests/vector_otlp_relay.rs`. The test pushes an OTLP span into a Vector source whose
-sink points at a dead collector (Vector buffers it), brings the mock collector online, then polls ~30s
-(`Fixed::from_millis(250).take(120)`) for the buffered span to be relayed. The captured Vector stderr shows no sink
-errors -- the buffer just wasn't flushed inside the window, so Vector's post-reconnect sink retry/backoff on a cold
-runner is the suspect. Observed once on commit `a6d7cd5552586d95dc67e09d5800c347f368a998` at
-https://github.com/edge-toolkit/core/actions/runs/28690968262/job/85092063432 and passed on the next run with no
-code change. If this signature recurs, stop rerunning and fix the root cause: widen the poll window, or configure
-the Vector sink's retry/backoff so a freshly-online collector is hit promptly rather than after a long backoff.
-
 ## Workarounds
 
 When you can't (or shouldn't) fix the root cause right now -- a libc race in an upstream dep, a flaky platform driver,
diff --git a/config/vector-otlp-relay.yaml b/config/vector-otlp-relay.yaml
index 1765ce6..f03ca68 100644
--- a/config/vector-otlp-relay.yaml
+++ b/config/vector-otlp-relay.yaml
@@ -35,7 +35,13 @@ sinks:
       # `otlp` is protobuf; Vector's json OTLP encoding is not a valid request body (vectordotdev/vector#23971).
       encoding:
         codec: otlp
-      # Cap the retry backoff so the buffered event is redelivered within seconds of the collector returning.
+      # Retry promptly but keep retrying long enough to survive an outage.
+      # retry_initial_backoff_secs makes the first retry quick, so a buffered event is redelivered within a
+      # second or two of the collector returning. retry_max_duration_secs is NOT a backoff-interval cap --
+      # it is the total per-request retry window, after which Vector DROPS the event ("retries exhausted").
+      # At 5s the buffered span was dropped before a cold CI runner could bind the collector, which timed out
+      # the ws-wasi-runner vector_otlp_relay test on Windows; 300s keeps store-and-forward durable across a
+      # real outage while staying finite.
       request:
         retry_initial_backoff_secs: 1
-        retry_max_duration_secs: 5
+        retry_max_duration_secs: 300