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

[dc0sk]

Nothing coordinates two heavy cargo runs on one machine. On 2026-09-25 a
neighbouring project's gate (fnec-rust) was OOM-killed three times while this
repo's cargo test --workspace ran alongside it — and every kill read as a
failure of the code under test, which is the misattribution GATE: INVALID
exists to prevent, arriving by a channel this gate cannot see.

scripts/lib/host-build-lock.sh takes a HOST-WIDE flock, held until the
calling script exits. scripts/gate.sh and the pre-push hook each source it
once. The path is shared with fnec-rust's scripts/host-build-lock.sh, so the
two projects' gates queue instead of contending for RAM; HEAVY_BUILD_LOCK
overrides it, and every project must agree.

Placement is deliberate:

• gate.sh takes it right after argument parsing and BEFORE START_HEAD, so
  time spent waiting is not part of the run and the drift guard's baseline is
  taken after the wait.
• --fingerprint is exempt — the seconds-long primitive builds nothing.
• the pre-push hook takes it AFTER the GATE_SKIP escape hatch, so a
  deliberate skip never waits.
• neither calls the other, so both may take it without deadlocking.

Verified without building (a build in this worktree would put a target dir
inside it): with a test lock held, scripts/gate.sh and the pre-push hook each
printed "waiting for it to finish" and were stopped by timeout before any
cargo step (exit 124); --fingerprint returned in 1.2 s against a 6 s hold.

NOT verified: this repo's rule to sabotage-verify gate.sh through a full run
after every edit. A full gate is ~2 h and this checkout's sibling was running
the ota_channel_adaptation acceptance suite at the time; running a second
suite beside it is the exact contention this change exists to stop. Run
scripts/gate.sh on a quiet machine before merging.

Verification-objective: two heavy cargo runs on one host serialize instead of
racing for RAM — observed as a gate and a pre-push hook each blocking on a
held host lock, and the exempt fingerprint mode not blocking

Pushed with GATE_SKIP=1, visibly and deliberately: the diff contains no Rust, the hook would have run workspace clippy inside a worktree, and the sibling checkout was mid-way through the ota_channel_adaptation acceptance suite. Do not merge until scripts/gate.sh has run on a quiet machine (this repo's own sabotage-after-every-edit rule).

The lock only protects a checkout once cargo-husky reinstalls the hook from .cargo-husky/hooks/pre-push — the installed copy in .git/hooks currently differs from the versioned source.

🤖 Generated with Claude Code

https://claude.ai/code/session_018p7FxX7QMWNJVNp9LbaLkB