[CI] Build the Windows and macOS legs without debug information (#7798)

* [CI] Build the Windows and macOS legs without debug information

Debug information is most of what a compiler cache stores: one Linux
configuration measures 1.2 GB with full DWARF and 660 MB at -g1. The Windows leg
cannot be lowered that way, because MSVC has no -g level and /Z7 is the only
cacheable form, so it sits at full debug info and fills a 1.2 GB cache.

Teach COCKATRICE_DEBUG_INFO_LEVEL=0 to mean something on MSVC by leaving /Z7
out, and build the vcpkg legs (Windows and macOS) at level 0 for every non-tag
ref. A tag is the build that ships and is symbolicated, so it keeps the full
set. The Linux legs keep -g1: a sanitizer report without file and line is not
worth the quota it would save.

Report the ccache directory and the object total at the end of each leg, so the
working set a bound has to fit is a number in the log rather than an estimate.

* [CI] Size the Windows cache for the no-debug footprint

Measured on run 42377451921, one Windows configuration without /Z7 writes a
890 MB ccache directory, against 1.2 GB with it. A 950M bound holds that with
5.4% headroom, which is the same rule the Linux bounds follow.

The comment block that splits the quota follows: 4.25 GB of bounds, 3.9 GB of
stored compiler cache, 8.1 GB of 10 in total.

.github/workflows/desktop-build.yml

@@ -248,10 +248,10 @@ jobs:
       #   Debian 13            Release + suite     625 MB               800M
       #   Fedora 45 debug      Debug + suite       660 MB               850M
       #   Fedora 45 release    Release + package   614 MB               800M
-      #   Windows 10           Release + package   1.2 GB working set   1300M
+      #   Windows 10           Release + package   890 MB working set   950M
       #
-      # That is 4.6 GB of bounds; what lands in the quota is 4.3 GB of compiler cache,
-      # 2.97 GB stored by the Linux legs and 1.30 GB by Windows, for 8.4 GB of 10 in total.
+      # That is 4.25 GB of bounds; what lands in the quota is 3.9 GB of compiler cache,
+      # 2.97 GB stored by the Linux legs and 0.95 GB by Windows, for 8.1 GB of 10 in total.
       # So only legs worth their share get one. A leg keeps no cache when it duplicates
       # another leg's coverage (Ubuntu 26.04 behind Debian 13's Release + suite; the macOS
       # legs behind the Linux ones) or finishes well inside the Windows leg (Servatrice
@@ -680,7 +680,7 @@ jobs:
             qt_modules: qtimageformats qtmultimedia qtshadertools qtwebsockets
             type: Release
             ccache_eviction_age: 7d
-            ccache_size: 1300M    # Release plus the package: 1.2 GB working set at /Z7
+            ccache_size: 950M    # Release plus the package: 890 MB working set without /Z7
             use_ccache: 1
             # Scope is named after the vs2026 toolchain, not the generator: the
             # ABI hash covers cl.exe, which is identical, so this leg keeps
@@ -702,8 +702,11 @@ jobs:
       CCACHE_DIR: ${{ github.workspace }}/.cache/
       CCACHE_SIZE: ${{ matrix.ccache_size || '700M' }}
       PERSIST_CCACHE: ${{ matrix.persist_ccache || 'true' }}
-      # MSVC has no -g level, so the Windows leg does not ask for one.
-      COCKATRICE_DEBUG_INFO_LEVEL: ${{ matrix.os != 'Windows' && github.ref_type != 'tag' && '1' || '' }}
+      # MSVC has no -g level, so there is nothing to ask for on Windows: level 0 is the one
+      # value with a meaning there, and it drops /Z7 rather than lowering a level. Pull-request
+      # builds carry no debug information, which is most of what a compiler cache stores; a
+      # tag is the build whose binary is shipped and symbolicated, so it keeps the full set.
+      COCKATRICE_DEBUG_INFO_LEVEL: ${{ github.ref_type != 'tag' && '0' || '' }}
 
     # Shared with build-vcpkg-nightly so the two leg sets run the same steps.
     steps: &vcpkg-steps
@@ -877,6 +880,21 @@ jobs:
           WITH_CRASH_REPORTING: ${{ matrix.type == 'Release' && 'ON' || '' }}
         run: .ci/compile.sh --server --test --vcpkg
 
+      # What the compiler cache has to hold for this leg. ccache stores the objects it
+      # compiles, so their total is the working set a bound has to fit, and the cache
+      # directory shows what this build actually wrote. Debug information dominates both,
+      # which is why the level above is reported next to them.
+      - name: "Report compiler cache footprint"
+        if: always()
+        shell: bash
+        run: |
+          echo "=== ccache directory"
+          du -sh "$CCACHE_DIR" 2>/dev/null || echo "(none)"
+          echo "=== objects and archives"
+          find build -type f \( -name '*.o' -o -name '*.obj' -o -name '*.a' -o -name '*.lib' \) -exec du -k {} + 2>/dev/null \
+            | awk '{n++; s+=$1} END {printf "%d files, %.2f GB\n", n, s/1048576}'
+          echo "=== debug info level in use: '${COCKATRICE_DEBUG_INFO_LEVEL}'"
+
       - name: "Install sentry-cli"
         if: needs.configure.outputs.tag != null && matrix.make_package == '1'
         # Same as the Linux leg above: the install downloads from Sentry's registry, so a