xlink and release-diff triage; finding 59 (VxWorks linker-flavor ICE)

[zmaril]

Results of the big run's xlink and release-diff, triaged.

• xlink harness fixes:
  • Windows MSVC targets with a custom entry point need /SUBSYSTEM:CONSOLE, which lld-link can't infer.
  • Symbols the platform's C runtime provides count as environment failures: the MSVC CRT under /NODEFAULTLIB, and the __atomic_* builtins NuttX supplies by design.
  • Architectures lld can't link count as environment failures too: csky, m68k, sparc, xtensa, ppc64 ELFv1, aarch64 ILP32.
  • Timeouts are their own class.
• release-diff: "slower" now compares CPU time. run_command reaps with wait4, so every Finished carries the CPU time of the command and its descendants.
  • The big run flagged dioxus and helix as about 1.9× slower. That was load noise: on helix both nightlies use 12.2 s vs 12.7 s of CPU.
• checks.md: rows for xlink (222 of 334 targets link) and scale-check.
• Finding 59: the 9 *-wrs-vxworks targets ICE (assertion failed: l.is_cc()) with -Clinker-flavor=ld.lld. Tier 3, low severity. It reproduces by hand; the facts doc has the command.

Numbering assumes #41 (56–57) and #42 (58) merge first.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QXiEXbESemwqMLYKaWLDbT