boot: STARTUP.EXE reads OVMXVMSSYS.PAR, sets real hostname from SCSNODE (vms-b6a7)

[baron-3dl]

Summary

• STARTUP.EXE's SYSBOOT role (src/ovmx_init/ovmx_init.c) now reads SYS$SYSTEM:OVMXVMSSYS.PAR after the system disk is mounted and confirmed installed, and calls sethostname() from the SCSNODE parameter. The hardcoded sethostname("OVMX", 4) is gone.
• Shares the existing reader (sysgen_read_string()/sysgen_read_param() in src/libvms/include/sysgen_params.h) — the same header vms_sysgen.c, scsd.c, and the DCL F$GETSYI lexicals already use. No second parser (vms-9b7 defect class). PID 1 needs no new link dependency: it already links vmsfs/vmsfs_static.
• Missing/corrupt OVMXVMSSYS.PAR on an already-installed disk is not the mount-or-halt condition: it boots on the compiled-in default node name with an honest, explicitly %OVMX--facility warning (Rule 10 — no invented VMS message, since no oracle capture of this condition exists).
• New QEMU e2e gate tests/qemu/test_boot_scsnode_hostname_e2e.sh (+ ctest entry + CI job boot-scsnode-hostname-e2e), proving against a real boot/reboot: SYSGEN SET SCSNODE + WRITE CURRENT survives a reboot into the real Linux hostname (via F$GETSYI("NODENAME"), not SCSNODE/SHOW SYSTEM which already read the .PAR directly and would pass unchanged); and the missing-.PAR negative control.

Harness hang (found and fixed during verification)

A local proof run stalled for 1h43m. Root cause, isolated by direct reproduction: launching QEMU through a helper function called via command substitution (QP=$(boot_qemu ...)) reliably wedged the backgrounded QEMU launch — ps inside the container showed only the stuck subshell, no timeout/qemu child at all, versus an inlined equivalent that boots to Username: in ~9s every time. Fixed by inlining every QEMU invocation directly, matching the already-proven pattern in test_release_e2e.sh / test_sysgen_versioning_e2e.sh — no boot_qemu() helper. Also hardened: timeout -k 15 (SIGKILL escalation if QEMU ignores SIGTERM) on every invocation, plus an outer timeout --kill-after=30 <budget> + named-container cleanup trap in run_boot_scsnode_hostname_e2e.sh so no combination of internal blocking calls can hang past a fixed ceiling.

Along the way, found and fixed two test-assertion bugs (not product bugs): check()'s grep -qF lacked --, so patterns starting with - (VMS's real continuation-line convention, e.g. -OVMX-I-NOPARAMS, ...) were silently parsed as grep options and never matched; and a DIRECTORY-finds-nothing assertion expected an invented %DIRECT- error line where the real, correct output for an exact filespec with zero matches is Total of 0 files, 0 blocks..

Test plan

• Baseline ctest -LE slow: 123/124 passing pre-change; the one failure (env_identity_census_negctl, a slow whole-tree-copy mutation test, unrelated to this change) reproduced identically on the unmodified tree — confirmed pre-existing, not introduced here.
• Full rebuild (host + Docker bootable image) compiles clean on top of current main (dcl: MOUNT/DISMOUNT really mount(2)/umount(2) vmsfs volumes (vms-651) #277 MOUNT, PRODUCT INSTALL: land a real kit on a real vmsfs volume (vms-df9) #278 PRODUCT INSTALL, release: cut-release.sh — reproducible release bundle from a clean tree (vms-d73) #279 cut-release, release: install->UPGRADE->boot e2e gate proves the upgrade-safety gap (vms-f05) #280 upgrade-gate).
• tests/qemu/test_boot_scsnode_hostname_e2e.sh against the real mastered image: 21/21 checks green (positive: SCSNODE survives reboot into the real hostname; negative: missing .PAR boots honestly on defaults).
• tests/qemu/test_persistent_boot.sh (mount-or-halt gate) against the same image: 20/20 checks green — confirms the identity change does not regress the mount-or-halt boot gate.
• CI will additionally run all existing boot/persistent-boot/release-e2e/executive-integral/distrib-boot jobs plus the new boot-scsnode-hostname-e2e job.

🤖 Generated with Claude Code