Lane/tree collision admission alternatives — 2026-08-10

Verdict

INCOMPLETE. The local packet executes fak’s in-process admission and a geometry-only baseline. DOS arbitration, GitHub Actions, Kubernetes Lease coordination, and etcd have zero measurements until their real stores/runtimes enforce the same lock state; #6144 tracks those witnesses. Mis-scoped capacity issue #6143 was closed and replaced.

Same-workload contract

Every arm receives two live leases and five requests: same-lane/disjoint-tree, cross-lane/overlapping-tree, exclusive lane, declared read-only, and self-lease renewal. The same taxonomy supplies canonical trees and exclusivity. Full witnesses report false allows/denies, admission quality, decision/acquisition latency and throughput, CPU/RSS/network/storage, contention, and total cost.

Arm	Class	Local availability	Honest result
fak native lane and tree admission	native	yes	correct five-verdict sequence
geometry-only tree overlap	tuned baseline	yes	incomplete semantics: misses lane and exclusivity policy
DOS arbitrate	first-class integration	no	separate real lease-store run required
GitHub Actions concurrency groups	external	no	zero measurements
Kubernetes Lease coordination	external	no	zero measurements
etcd concurrency mutex	external	no	zero measurements

Mocks, parsers, and adapters are not distributed-lock witnesses.

Local native witness

go test ./internal/laneadmit -bench BenchmarkDecideLaneTreeAdmission -benchmem -run '^$' -count=5

Windows/amd64, AMD Ryzen 9 9950X: 5626, 6056, 6488, 7209, 9005 ns/op. Median: 6488 ns/five decisions (about 1298 ns/decision), 4551 B/op, 196 allocs/op. This is in-process policy cost, not distributed acquisition latency.

Reproduce

go test ./internal/laneadmit -run TestCompareLocalKeepsLockAlternativesExplicit
go test ./internal/laneadmit -bench BenchmarkDecideLaneTreeAdmission -benchmem -run '^$' -count=5
go test ./internal/nativebench