perf(ci): fine-shard ubuntu only; run whole suite on macos/windows

macOS GitHub runners have ~5 concurrent slots, so fanning 22 shards × 2 nodes
across macos/windows queued ~9 batches and pushed total wall to ~18min even
though each job was ≤60s — the queue depth, not per-job time, was the wall.

The ≤60s target is ubuntu-only (high runner concurrency). Restructure the test
matrix:
- ubuntu-latest: the 22 fine shards × node 22/24 (44 jobs) — all run in
  parallel, each ≤60s.
- macos-latest / windows-latest: whole suite as one job per node (shard=all,
  4 jobs) for compatibility coverage without flooding the scarce queue. This
  matches the pre-optimization per-OS job shape; flakes absorbed by --retry.

Drops test jobs from 132 → 48.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: killagu

.github/workflows/ci.yml

@@ -73,12 +73,13 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        os: ['ubuntu-latest', 'macos-latest', 'windows-latest']
+        # ubuntu-latest runs the fine-grained shards: fork-heavy packages are
+        # split into vitest --shard slices and the light remainder into
+        # weight-balanced rest-* shards (see scripts/run-shard.js). Sized so each
+        # shard's `vitest run` step stays ≤ ~60s on ubuntu-latest, where runner
+        # concurrency is high enough to run them all in parallel.
+        os: ['ubuntu-latest']
         node: ['22', '24']
-        # Fork-heavy packages are split into multiple vitest --shard slices and
-        # the light remainder into rest-* shards (see scripts/run-shard.js).
-        # Sized so each shard's `vitest run` step stays ≤ ~60s on ubuntu-latest;
-        # slower macos/windows runners may exceed that but still parallelize.
         shard:
           - cluster-1
           - cluster-2
@@ -102,6 +103,23 @@ jobs:
           - rest-4
           - rest-5
           - rest-6
+        include:
+          # macOS/Windows runners have very low concurrency (~5 macOS slots), so
+          # fanning out 22 shards × 2 nodes would queue ~9 batches and blow up
+          # wall time. Run the whole suite as a single job per (os, node) instead
+          # — coverage without flooding the queue. (shard=all => no filter.)
+          - os: macos-latest
+            node: '22'
+            shard: all
+          - os: macos-latest
+            node: '24'
+            shard: all
+          - os: windows-latest
+            node: '22'
+            shard: all
+          - os: windows-latest
+            node: '24'
+            shard: all
 
     name: Test (${{ matrix.os }}, ${{ matrix.node }}, ${{ matrix.shard }})
     runs-on: ${{ matrix.os }}