[Sunrisepeak]

Design: .agents/docs/2026-11-14-ci-and-release-time-cost-design.md §6 (round two; the owner confirmed W1–W4, the division of a case by the cases it declares, and the macOS paths on 2026-11-15).

Measured (PR run 43602794812, 97.6 min wall, 265 runner-minutes queued)

Interval (min)	Cause
0 → 9.1	the cancelled previous run's target-matrix / coverage (if: always()) queued 8.5 min for a runner and held the concurrency group
9.1 → 20.7	the first jobs queued 11.4 min: three full CI runs in the organization at once
20.7 → 69.3	the archive stage: 9.3 queued, 39.4 built (the tests and the library at -O2: this package's default profile is release)
69.3 → 97.4	the shards: 28.0 / 27.9 / 12.4 / 9.8 min; e2e/openkal/trees 1373 s and sets 1375 s ran their cases in one process each, the longest case 564 s

The macOS unit step took 31 min and the macOS matrix scan 29.4 min (59 cells one after another).

Changes

• W1, main's test build reused. main's archive stage keeps its build directory, with the blob of each tracked file (.github/tools/source_mtimes.py record), as the artifact mcpp-unit-build-<host>; a pull request's archive stage restores it and gives each unchanged file a time older than every output, so only what changed is built again. An artifact and not the cache: the repository's cache holds 9.98 of its 10 GB, and a build directory there would evict the hosts' homes.
• W2, [profile.ci] (-O0, no debug information, strip) for the tests on Linux (archive), macOS and Windows (mcpp-unit input profile): the library and its unit tests took 8988 s of compiler time at the default release profile and 4183 s with it.
• W3, a long case divided by the cases it declared (SPEC-017 O10 and P4.3, mtest 0.9.0, 0.9.0: a selection excludes and selects where declared (SPEC-017 P4.3), so a runner divides a long case by the cases it declares mtest#24). When a program's selection states except and optional, a listed case longer than its share with two or more measured declared cases is divided by the innermost of them: each part but the last selects its cases with ?<id> (a case since removed is not missing), the last selects the case without them (!<id>), so an unmeasured case runs once. A divided case's estimate is at least the sum of its declared cases. The store's case durations are kept by program and id (layers and sets exist in two programs each).
• W3 measured on CI, and the rule that follows. Divided unconditionally, e2e/openkal/sets (1375 s whole) took 909, 1109 and 1306 s as three parts, and trees (1373 s) 1070 and 1247 s as two: the cases of one case share what the first fills (toolchains, packages, build caches), and every part pays it again on its runner. A division is now kept only while it pays: each part records the case it divides (divides), the store keeps the case's newest whole duration and its newest division's longest part, a case is divided once to be measured so and then only while its longest part took at most four fifths of the case whole, and a case is estimated by its whole run (test_schedule::division_pays, test TestSchedule/ADivisionIsKeptWhileItPays). SPEC-017 O10, docs/08 and design 6.3 state it.
• W4: coverage runs if: ${{ !cancelled() }}.
• Matrix scan side by side: scan.sh runs each cell in its own directory, MATRIX_JOBS at a time (the processors), and prints in the table's order.
• The built-in platform is openkal-platform 0.28.0 (was 0.27.1): the group on openkal 0.25.3, and the rows i686-linux-musl, i686-openkal, x86_64-uefi and i686-dos-musl. Found by the clean-room check: openkal 0.25.3's example built natively and was refused for x86_64-openkal, because 0.27.1's group holds openkal 0.25.0. The e2e case 1025 reads the pin from builtin.toml; the target-row snapshot (10 lines) and the matrix's runtime cells (openkal-llvm-runtime@0.20.3) follow.
• mtest 0.9.0 for this package's tests and for the scaffold of mcpp new (kMtestRelease), so the division applies to them.
• Fixed, found by the division's parallelism: an install for a project takes the home's state turn (xlings::state_dir). It saves the home's environment (subos/default/.xlings.json) as well as the project's, and with the turn in the project's directory it collided on Windows with another project's install (e2e images/1033, run 43632995462). Test XlingsState/AProjectsInstallTakesTheHomesTurn.
• Fixed, found by the matrix scan's cells side by side: a host tool built by two builds of one home at once is published by both. The publish goes through mcpp.build.stage::stage_file (a temporary of each writer's own, an equal destination counted as published); with one temporary name for every writer, one removed the other's before its rename (publishing host tool 'openkal-tools:okld' failed: No such file or directory, macos-arm64, run 43636471152). Staging temporaries carry a random part as well as the clock. Test BuildStage/WritersOfOneDestinationDoNotCollide.
• CI: the Windows legacy code-page check runs its two cases with the profile the job built the tests with (MCPP_TEST_PROFILE=ci); under the default profile it built every test again and passed its 600 s timeout.
• Test: ConcurrentChildren/AReaderReturnsWhenItsOwnChildExits starts both programs once before its rounds; a first start on a fresh Windows runner outlasted the slow child (round 0 of run 43626158365, -427 ms, where a shared pipe returns both together).
• Docs: SPEC-017 2.3 (P4.3, O10), docs/08 (en, zh); CHANGELOG; version 2026.11.15.1.

Local: TestSchedule/ACaseIsDividedByTheCasesItDeclared and the scheduling tests; test_source_mtimes.py; a demo of four declared cases 8.2 s → 3.0 s, a renamed case run once in the last part and its old name not missing; W1 restore with no change 0.4 s, one interface module changed 88 s (cold 248 s).