PRE-PUSH-LATENCY: guard-cargo 6223 s cold in pre_push_local.sh — keyed shared target dir + overlap guard-tree #5983

Five-whys, from aprender-pace measurements (inbox 19:04Z) and scripts/pre_push_local.sh on branch pace/pre-push-local @c910afba93:

1. Why did the pre-push take >100 min? guard-cargo alone ran 6223 s cold (guard-tree ~11 min, quick tier 166 s).
2. Why cold? step_guard_cargo unsets CARGO_TARGET_DIR on purpose (match the runner env), so cargo builds in the worktree-local ./target; every fresh worktree starts empty.
3. Why also slow to start? On lambda ~/.cargo/registry is a symlink, so CARGO_HOME falls back to /cargo-home and the first run downloads crates.
4. Why 6223 s and not ~1/3 of that? Load average ~43 (74 sessions) on lambda; cargo competes with the fleet. Not measured: per-phase split (compile vs download vs wait).
5. Why serial? preflight, guard-tree, guard-cargo, quick tier run one after another; guard-tree (no cargo) could overlap guard-cargo.
   Root cause: no reuse across pushes. The one step that must see the runner env has no stable, keyed build cache.

Countermeasure (one): give guard-cargo a stable CARGO_TARGET_DIR + CARGO_HOME keyed by sha(Cargo.lock + rust-toolchain.toml), e.g. /mnt/nvme-raid0/targets/pre-push-guard-cargo-, and start guard-tree in parallel with it. Acceptance: warm run <= 15 min, AND the host rebuilds <= the CI unit count after the container build (the 157-vs-12 check), measured and posted; cold run unchanged.
Owner: aprender-pace. Parent: #5923. Requested by aprender-57.