vms-b9e: spawn_users SIGPIPE harness fix + Alpha H1/H2 SCALE diagnostic

[baron-3dl]

Summary

Two changes on this branch, DIFFERENT landing status:

1. KEEPER — tests/qemu/test_syssvc_spawn_users.c: adds signal(SIGPIPE, SIG_IGN), copying the guard test_syssvc_showterm.c already has. Without it, a broken pipe to an already-exited DCL.EXE session child SIGKILLs the whole suite mid-run, hiding every assertion after that point. With the signal ignored, the same condition falls into a checked write() branch and reports a named failure instead of a dead test. Test-harness fix only, no product-code change.

2. DIAGNOSTIC, NOT for merge as-is — tools/cross-alpha/test-init-syssvc-alpha.c: bumps OVMX_TEST_DEADLINE_SCALE 8→20 (the tcg_deadline.h clamp ceiling) and SUITE_TIMEOUT_SECS 300→700 in lockstep, to disambiguate rd vms-b9e's H1 (calibration — the 4 DCL suites blow their own inner poll deadlines under Alpha-TCG cold start, not the outer watchdog) from H2 (a real registration race in spawn_users that a bigger timeout cannot fix). CI run 34255023889 (workflow_dispatch on this branch) is currently exercising the Alpha vms-341 leg with this diagnostic. If the 4 DCL suites go green at SCALE=20, this raw bump gets replaced with a MEASURED Alpha-TCG cold-start budget + justification before merge (never a silent bump to turn red green, per INV-6). If spawn_users is still red, this diagnostic gets reverted and the real fix is the registration race itself.

H2 (spawn_users registration race) — x86_64 stress probe, cheap and decisive

Built a minimal custom initramfs from the ovmx-ktest image (same vms.ko + DCL.EXE + test_syssvc_spawn_users binary), booted qemu-system-x86_64 -accel kvm directly, and looped test_syssvc_spawn_users in-guest 500 times: 500/500 exit 0, 0/500 showed the "SPAWNed subprocess is registered in the executive process table" FAIL. The race does not reproduce on native x86_64/KVM after 500 real SPAWN+register+procscan cycles — rules out an always-on, arch-independent defect reachable at native speed. Does not rule out Alpha-TCG slowness widening a window x86_64 never opens; that needs Alpha-side gdb, tracked on vms-b9e.

Test plan

• tests/qemu/test_syssvc_spawn_users.c SIGPIPE guard: unaffected on x86_64 (harness change only, no product-code path changed)
• x86_64 stress probe: 500/500 clean (see above)
• CI run 34255023889 Alpha leg (vms-341, ~45-60min, backlogged/workflow_dispatch-only) — pending, will determine H1 vs H2 disposition for the diagnostic commit

rd item: vms-b9e

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_018q4jvj6GCkdbYjRvpy7Vwu

🤖 Generated with Claude Code

[baron-3dl]

DO-NOT-MERGE / DIAGNOSTIC (vms-b9e). This branch carries a landable SIGPIPE harness fix + a NON-shippable SCALE 8->20 diagnostic to disambiguate H1 (calibration) vs H2 (race) on the Alpha leg. x86_64 500/500 clean already ruled out a native race. Hold until Alpha run 34255023889 lands: then either (H1) replace the raw SCALE bump with a MEASURED cold-start budget, or (H2) revert the bump — keeping the SIGPIPE fix either way. Converted to draft so it can't be auto-reaped.