vms-98c: lib$spawn spawns a REAL DCL subprocess (exec-drive prereq A)

[baron-3dl]

What

lib$spawn was a Rule-9 / INV-6 facade: it fork+exec'd /bin/sh -c <command> — the Unix Bourne shell, not a DCL command interpreter — and returned SS$_NORMAL for a SHOW TIME that no DCL ever saw. This replaces it with a genuine DCL subprocess.

Prereq A of the vms-ec70 self-host spine #4 exec-drive (making MMK.EXE actually drive compile/link).

How the DCL child really runs the command

• Resolve SYS$SYSTEM:DCL.EXE through the VMS filespec translator (vmsfs_to_linux_path) — the same CLI image JOB_CONTROL and PROVISION exec to hand a user a session, so a redefined SYS$SYSTEM is honored.
• fork(), redirect SYS$INPUT/SYS$OUTPUT, then execl(dcl, "DCL", "-c", command) — DCL executes the one command and exits (dcl_main.c -c mode). No command → interactive DCL on SYS$INPUT.
• CLI$M_NOWAIT honored (create-and-return); wait mode HIBERNATEs on the child and returns its completion status.

Honest boundary (Rule 9 / INV-6)

If the CLI image cannot be resolved or is not an executable regular file, lib$spawn returns an authentic SS$_NOSUCHFILE and runs nothing — no /bin/sh fallback, never a fake success for a command that did not run. A real child image needs no /dev/vms.

Test (tests/libvms/test_lib_spawn_dcl.c)

Redirects SYS$SYSTEM to a private temp root, stages the build's own DCL image there as DCL.EXE, and drives the real lib$spawn resolve+fork+exec path (no monkeypatch, hermetic, no shared /vms, no /dev/vms):

1. Observable effect from a real DCL child — lib$spawn("SHOW TIME") with SYS$OUTPUT redirected; asserts the captured file carries the current year (oracle: this process's own clock). Captured 13-AUG-2026 06:13:44.05 — genuine DCL SHOW TIME. /bin/sh -c "SHOW TIME" would print a SHOW: not found error, so the pre-fix body fails this.
2. Honest failure, no fallback — with the CLI image absent, asserts even (failure) status and that no output file was created.

vmsdcl -c runs without the executive (cf. test_dcl_basic.sh), so the whole spawn→DCL→SHOW TIME path is provable on a host ctest.

Results

• test_libvms_lib_spawn_dcl PASS; 34/34 libvms tests green.
• Rule-9 runtime_target_gate + runtime_target_negctl + dcl-integration green.
• Full tree builds clean.
• No new cross-image symbol (lib$spawn already exported; helper is static) → libvms_shr vector unchanged, native-link unaffected.

Deferred (prereqs B/C — vms-e0b mailbox, vms-9003 write-attention AST)

The efn/AST completion notification, named-subprocess ($CREPRC) registration, and the persistent DCL subprocess + mailbox + write-attention-AST protocol MMK's build_target.c uses (stream many commands into one long-lived DCL, read each command's full $STATUS from an end-of-command marker) are accepted-and-documented, not faked. This PR is the create+run primitive that rides under that protocol. Note: DCL's -c mode collapses $STATUS to shell 0/1, so full $STATUS fidelity comes via the B/C mailbox EOM markers, not this path.

Clean-room (Rule 8): LIB$SPAWN semantics from the public VSI OpenVMS RTL Library (LIB$) Routines Reference Manual and the DCL Dictionary (SPAWN / SHOW TIME).

🤖 Generated with Claude Code