[Sunrisepeak]

Summary

• K11. execvp, execvpe, execlp and posix_spawnp search the PATH through one function (port/src/okm_spawn.c). It takes the list separator and the program suffixes from kal_env_system (openkal clause 11 entry 30). Where the system is windows: PATH is split on ;, a name holding a backslash or a volume is not searched for, and in each entry a name that carries a suffix is tried as given and one without it only with each PATHEXT suffix in order (.COM;.EXE;.BAT;.CMD where unset), as Windows' command interpreter searches. Other systems keep POSIX's colon and no suffix. musl's execvp.c is excluded, so both ways of searching agree. The macOS cross-build scripts skip it too.
• K5. link and linkat with flags 0 run over kal_fs_hard_link. The errors: a directory → EPERM; two volumes → EXDEV; a volume without KAL_FS_PROP_HARD_LINKS → ENOTSUP; a taken name → EEXIST; any flag → EINVAL. The function is weak, so an image run by an older loader answers as before.
• E14. requires = ["mcpp:kernel-abi=openkal"] states the arrangement.
• Requirements move to the 0.25 family: openkal ^0.25, linux ^0.26, macos ^0.23, windows ^0.21, opensbi ^0.9 (^0.8.2 requires openkal 0.16). Version 0.29.0. README and musl/PATCHES.md are updated.
• CI: okrun is built with mcpp 2026.11.9.1, the floor openkal-run states. Unpinned, the step received 2026.10.24.1 and failed.

Tests: the probes and the bench speak the mtest protocol

tests/mtest_protocol.h is a header-only C writer of the mtest protocol, version 1, translated from openkal's conformance writer:

• P2: the handshake through MTEST_PROTOCOL/MTEST_REPORT and MTEST_MODE=list, with the mtest-protocol=1; marker.
• P3.4: one case per probe, one part per observation.
• P4: MTEST_SELECT, with lines that name nothing reported as missing.
• P5: one write per event.
• P6: the bench reports one bench case per measured primitive.
• P7.1: a serial resource.

A copy a probe starts does not inherit the runner's variables.

They are protocol programs and not mtest UTest/Bench objects because of layering, not preference. mtest's library is C++ that imports std. The only C++ runtime over this C library is openkal-llvm-runtime's libc++, which is built above it. A spike in this package's own graph fails: "std module precompile failed: … fatal error: features.h: No such file or directory", with c++-abi "—". The UTest form of what runs above this library is in openkal-llvm-runtime's tests (mcpplibs/openkal-llvm-runtime#70). That is also where the Windows PATH/PATHEXT observation of hostname runs: this repository's CI has no Windows row.

The C writer is a candidate for mtest itself, as a C adapter for any C project.

CI: tools/run-probe.sh keeps its text verdict and runs each probe a second time under mtest run when MTEST is set. The bench runs once under mtest. Each job writes mtest report --format md to its summary and keeps the store as an artifact, and a report job joins every job's runs.

Test plan

• Existing suite at the functional change: CI run 42709051859 green at dd35bef (the posix probe extended with link/linkat and with posix_spawnp past an absent PATH entry).
• Protocol migration: CI run 42713947761 green at 0671d8f, all 11 checks. Each of the 19 probes run by tools/run-probe.sh also passes under the runner on the Linux gcc, Linux llvm and macOS jobs (posix: 82 passed). The bench under the runner reports 69 cases (68 on macOS). The tests (report) job joins them.
• Windows suffix order (review): CI run 42717890519 green at b9e95e0, all 11 checks, both suites.