vms-6a2: give the corpus conformance gate real teeth (floor 90/229, vacuity guard, path fix)

[baron-3dl]

Problem (vms-6a2)

The corpus conformance gate was documented as a falsifiable floor but had real gaps:

1. Vacuity risk — a total=0 baseline/report makes the whole gate exit 0 forever (an empty floor can never be undercut). This is the original toothless-gate defect the item was filed on; nothing mechanically prevented the baseline reverting to total=0.
2. Container-path hardcoding — run_corpus.sh hardcoded /src/... for includes, -L, and LD_LIBRARY_PATH, so it only ran correctly inside the CI docker mount, not from a plain host checkout/worktree.
3. Flaky boundary count — the run-pass count flapped (90 vs 86 across back-to-back runs), so a floor set at the high value would redden CI intermittently.

What changed

tests/conformance/run_corpus.sh

• Derive REPO_ROOT from SCRIPT_DIR (= /src in CI, unchanged) instead of hardcoding /src; adds OVMX_BUILD_LIB_DIR override. Runs from any checkout/worktree now.
• Redirect program stdin from /dev/null. The input routines (lib$get_input/get_command/get_foreign/lookup_key) flapped run-pass↔run-fail on an inherited tty. /dev/null = the closed stdin CI already gives, so the measurement is deterministic (root-cause fix, not a retry/timeout band-aid).
• Vacuity guard: hard-fail (exit 2) when total==0.

.github/workflows/ci.yml

• Floor step rejects total==0 on the report or the committed baseline (independent backstop to run_corpus.sh's exit 2).

tests/conformance/corpus_baseline.json

• Real measurement against current HEAD libs, floor ratcheted 89 → 90 run-pass / 229 total (compile-fail 66, link-fail 52, run-fail 7, run-crash 14). Timestamped, per-program status. Stable over 5 runs after the stdin fix.

Real measured numbers (current HEAD, x86_64 workshop)

metric	value
total	229
compile-fail	66
link-fail	52
run-pass (the conformance number)	90 (39%)
run-fail	7
run-crash	14
compile-or-better	111 / 229 (48%)

Note for the record: the closed vms-801.4 "80%+ compile and run" claim is not backed by the tree — real run-pass is 39%, compile-or-better 48%. The committed baseline now carries the honest number either way.

Proof the gate has teeth (all foreground, exit codes read directly)

control	result
current state (positive)	run_corpus.sh exit 0, floor held 90 ≥ 90
per-entry regression	run_corpus.sh exit 1 (faked a baseline run-pass entry → flagged lib_attach)
aggregate floor	CI step fails (current 90 < simulated floor 91)
total==0 (report)	run_corpus.sh exit 2 + CI floor step exit 1
total==0 (baseline)	CI floor step exit 1

Runs clean from a fresh worktree checkout (proved here — the worktree is not at /src), and CI behavior is unchanged since REPO_ROOT resolves to /src there.

🤖 Generated with Claude Code