chore(gates): queue behind other heavy builds on the host, not race them

[dc0sk]

Why

Nothing coordinates two heavy cargo runs on one machine. On 2026-09-25 this repo's gate was OOM-killed three times while another project's cargo test --workspace (OpenPulseHF) was running at the same time. Each kill looked like a failure of the code under test.

What

scripts/host-build-lock.sh takes a host-wide flock and holds it until the calling script exits. scripts/check-all.sh and .githooks/pre-push each source it once, near the top.

The lock path is shared with OpenPulseHF's scripts/lib/host-build-lock.sh (dc0sk/OpenPulseHF#1440), so the two projects' gates queue instead of competing for RAM. HEAVY_BUILD_LOCK overrides the path, and every project has to use the same value.

There's one helper, sourced by both scripts. Pasting the snippet into each would create two copies of the same decision that can drift apart, which is this repo's most common defect.

Deadlock rules (stated in the helper):

• A script takes the lock once, and nothing it calls takes it again. pre-push does not call check-all.sh.
• Nobody wraps a script that takes the lock in flock by hand. The inner take would wait forever on the lock the outer one already holds.

Verification

• Contended: with the lock held by another process, a second take printed "waiting for it to finish". It got the lock 3.5 s later, exactly when the holder released it.
• Uncontended: it got the lock in 0.01 s.
• This PR's own push: the pre-push hook took the lock and ran the full suite, after I'd waited for OpenPulseHF's gate to finish (it did at 11:36).

Context

The same session also:

• replaced the "don't run two at once" advice in the global CLAUDE.md with a rule that requires this lock;
• added one line to the code-quality-gates skill's pre-push template.

🤖 Generated with Claude Code

https://claude.ai/code/session_018p7FxX7QMWNJVNp9LbaLkB