ci: four Linux shards from the measured durations; e2e 741 compiles a library and links nothing

Author: speak-agent

.agents/docs/2026-10-02-pr-ci-acceleration-and-the-toolchain-specification-design.md

@@ -564,6 +564,74 @@ measurement uses the same scripts as §2.1.
 - Cache usage below 8 GB.
 - A documentation-only pull request finishing under 5 minutes.
 
+### 2.7 What was built, and where it departs from §2.4
+
+Implemented in mcpp#759. Each departure below was decided by a measurement or a
+reading taken while building it.
+
+- **The artifact is the binary itself, not a packed copy (§2.4.1).** The Linux
+  self-host binary has the interpreter
+  `~/.mcpp/registry/data/xpkgs/xim-x-glibc/2.44/lib64/ld-linux-x86-64.so.2`, needs
+  `libgcc_s.so.1` from the `xim-x-gcc/16.1.0` payload, and links libstdc++
+  statically (measured locally, `readelf`). A consumer that restores the
+  sandbox runs it as it is. Packing it would test a different binary from the
+  one every self-host build produces. `use-built-mcpp` runs the binary first,
+  and only when it does not run does the bootstrap install the toolchain
+  `mcpp.toml` names for the host; a binary that still does not run fails the
+  step. The binary is 25 MB.
+- **The Wine packages keep their cache (§2.4.4).** Their eviction was a
+  consequence of F3, not a property of the cache. With one writer per key, the
+  `wine-debs` cache is saved only by `mingw-cross-wine` on main and is no longer
+  displaced. A release asset would have added a second thing to publish and
+  keep current.
+- **The timing tables are in the repository (§2.4.5).** They live under
+  `tests/e2e/timings/<host>.tsv`, seeded from the per-test lines of the
+  2026-10-01 logs. A shard's membership is then a function of the commit,
+  which makes it reproducible (`E2E_LIST=1` prints it). The `e2e-coverage` job
+  uploads the merged durations of each run as the artifact `e2e-timings`, and
+  refreshing a table is copying a file. The tables were seeded from the
+  2026-10-01 logs, then replaced by the durations of the first run of this
+  change, in which 515 Linux tests ran instead of 466. Linux therefore has four
+  shards rather than three, each budgeted at 9.9 minutes. Windows has three at
+  14.4 to 14.5, and macOS two at 9.1 and 9.3. The step limits are about twice
+  the budgets: 22, 30 and 20 minutes.
+- **The classifier searches exact paths (§2.4.3).** It searches for the
+  changed path and for its translation (`docs/X` and `docs/zh/X`), and not for
+  a bare file name. A search on the name made `.agents/docs/README.md`, which
+  every new record regenerates, a code change, because release packaging names
+  `README.md`. The checks of the `docs` job do not count as readers, because
+  they run on every change. A document that any other script or source names
+  starts the whole CI rather than a subset of it, which is simpler and errs
+  towards running more.
+- **The macOS legs of the target matrix and of openkal still build (§2.4.1).**
+  They run on `macos-14`, and the artifact is built on `macos-15`. release.yml
+  records that a bootstrap build linking the system libc++ with a minimum
+  version of 14 failed at launch on `macos-14`. Until the artifact is measured
+  there, those two legs build their own, and the gate carries
+  `ci-lint: allow-r1` with that reason. Their Linux and Windows legs consume the
+  artifact.
+- **`invariants` and `scan` remain two jobs (§2.4.1).** Merging them is a
+  rewrite of the target-matrix workflow beside a change that already rewrites
+  twelve; it is left for a change of its own.
+- **A test that never ran was broken.** 741, one of the seven `llvm` tests,
+  failed on its first run in CI. The root package of its fixture was a binary,
+  so the build linked `cabi-probe.exe` for `x86_64-windows-gnu`, and with
+  `allow_host_libs` the link found the host's mingw-w64 libraries. It passed on a
+  machine that has them and failed on every runner, and nothing noticed, because
+  no runner ran it. The fixture's root is now a library, which is what the test
+  says it is: "this test compiles only".
+- **Coverage found more than F9.** Classifying the tests of the 2026-10-01
+  logs found 24 that ran on no runner and were named by no workflow. The
+  seven `llvm` tests are among them, and so are three `musl` tests: the probe
+  named 15.1.0 while the runners installed 16.1.0. Seven `mingw-cross` tests
+  are in the list too, a toolchain no shard installed. The rest were 105 (nasm),
+  65 (scan-deps), 239 (named by its `E2E_ONLY` pattern, which the check now
+  reads), 257 (needs wine and a Linux-hosted MinGW), 658 (an attached Android
+  device), and 873-877 (added after the logs). The capability probes now ask for
+  a family. The Linux shards install musl, llvm, mingw-cross and nasm. 257 runs
+  in `mingw-cross-wine`. 658 is the one entry of
+  `tests/e2e/coverage-exceptions.tsv`.
+
 ## Part III. Defects repaired by the next release
 
 ### 3.1 #757: the engine's identity is part of a build record

.github/workflows/ci-linux-e2e.yml

@@ -1,7 +1,7 @@
 name: ci-linux-e2e
 
 # The e2e suite (tests/e2e/run_all.sh) on Linux, called by ci.yml beside
-# ci-linux.yml after the Linux build. Three shards, assigned by measured
+# ci-linux.yml after the Linux build. Four shards, assigned by measured
 # duration (tests/e2e/timings/linux.tsv, rule R4 of the 2026-10-02 CI record),
 # each running the one binary build.yml produced (use-built-mcpp, rule R1).
 # Each shard writes a per-test report that the e2e-coverage job of ci.yml reads
@@ -15,15 +15,15 @@ on:
 
 jobs:
   e2e:
-    name: e2e ${{ matrix.shard }}/3 (linux x86_64, self-host)
+    name: e2e ${{ matrix.shard }}/4 (linux x86_64, self-host)
     runs-on: ubuntu-24.04
     timeout-minutes: 45
     strategy:
       fail-fast: false
       matrix:
-        shard: [1, 2, 3]
+        shard: [1, 2, 3, 4]
     env:
-      E2E_SHARD: ${{ matrix.shard }}/3
+      E2E_SHARD: ${{ matrix.shard }}/4
       E2E_TIMINGS: tests/e2e/timings/linux.tsv
       E2E_REPORT: ${{ github.workspace }}/e2e-report-linux-${{ matrix.shard }}.tsv
       MCPP_HOME: /home/runner/.mcpp
@@ -49,9 +49,9 @@ jobs:
           host: linux-x86_64
 
       - name: E2E suite
-        # Twice the shard's budget of about ten minutes (R4): reached only by a
+        # About twice the shard's budget of ten minutes (R4): reached only by a
         # hang. The per-test 600 s limit in run_all.sh names the test that hung.
-        timeout-minutes: 25
+        timeout-minutes: 22
         run: |
           set -euo pipefail
           # MCPP is this commit's binary and MCPP_BOOT the released bootstrap

.github/workflows/ci-windows-e2e.yml

@@ -81,15 +81,13 @@ jobs:
 
       - name: E2E suite
         shell: bash
-        # Fail-fast on hung tests instead of burning the whole job budget.
-        # Per-test 600s timeout lives in run_all.sh.
-        timeout-minutes: 25
+        # About twice the shard's budget of fifteen minutes (R4): reached only by
+        # a hang. The per-test 600 s limit in run_all.sh names the test that hung.
+        timeout-minutes: 30
         run: |
-          # The RELEASED mcpp this job bootstrapped from, kept for e2e 252: the
-          # claim that an older client can still build against a package the PR
-          # produces is only worth making if something checks it against a real
-          # old binary. Captured before $MCPP is repointed at the fresh build.
-          # MCPP_BOOT is the released bootstrap (use-built-mcpp), for e2e 252.
+          # MCPP_BOOT is the released bootstrap (use-built-mcpp). e2e 252 checks
+          # an older client against a package this commit produces, so it needs
+          # a real old binary.
           export MCPP_BOOT
           export MCPP="$MCPP_SELF"
           export MCPP_VENDORED_XLINGS="$XLINGS_BIN"

tests/e2e/741_c_abi_environment_declared_by_the_c_library.sh

@@ -63,6 +63,12 @@ cabi_probe_asm_marker:
     ret
 EOF
 
+# THE ROOT IS A LIBRARY, SO NOTHING LINKS. With the inferred binary target the
+# build linked `cabi-probe.exe` for x86_64-windows-gnu, which needs a MinGW
+# runtime: it passed on a machine with the host's mingw-w64 installed and failed
+# on every CI runner, and nothing noticed because no runner ever ran this test
+# (it declared `llvm`, which no runner granted, until 2026-10-02). Its subject
+# is the compile lines, which a library has as well.
 cat > mcpp.toml <<'EOF'
 [package]
 name    = "cabi-probe"
@@ -72,6 +78,10 @@ version = "0.1.0"
 fakemusl  = { path = "fakemusl" }
 openkalwin = { path = "openkalwin" }
 
+[targets.cabi-probe]
+kind    = "lib"
+sources = ["src/*.cpp", "src/*.S"]
+
 [build]
 allow_host_libs = true
 EOF