Shared region admission alternatives — 2026-08-10

Verdict

INCOMPLETE. The local packet executes fak’s shared execution-surface admission and a tuned geometry-only incumbent. DOS arbitration, Git-ref lease acquisition, Kubernetes, etcd, and GitHub Actions retain zero measurements until their real stores and runtimes enforce the same region state; #6149 tracks those witnesses.

internal/regionadmit is distinct from the lower-level internal/laneadmit contract: it resolves or infers regions at the execution-surface seam, permits two strictly narrowed disjoint sub-regions, and consumes lease projections from multiple acquisition systems. Those acquisition systems remain separate integration arms rather than being counted as native evidence.

Same-workload contract

Every arm receives two live leases, one taxonomy, and six requests: narrowed same-lane/disjoint-region, cross-lane/overlapping-region, exclusive lane, declared read-only, self-renewal, and hierarchical sub-lane. Full witnesses report false allows/denies, admission precision/recall, decision and acquisition latency, throughput, CPU/RSS/network/storage, and total cost.

Arm	Class	Local availability	Honest result
fak native shared region admission	native	yes	correct six-verdict sequence
geometry-only region overlap	tuned baseline	yes	incomplete: misses exclusivity and hierarchical lane serialization
fak + DOS arbitrate	first-class integration	no	real arbiter and lease-store run required
fak + Git-ref leases	first-class integration	no	real cross-process acquisition and visibility run required
Kubernetes Lease coordination	external	no	zero measurements
etcd concurrency mutex	external	no	zero measurements
GitHub Actions concurrency groups	external	no	zero measurements

Adapters, parsers, and in-memory stand-ins are not distributed acquisition witnesses.

Local native witness

go test ./internal/regionadmit -bench BenchmarkDecideSharedRegionAdmission -benchmem -run '^$' -count=5

Windows/amd64, AMD Ryzen 9 9950X: 8068, 8089, 10218, 9977, 12329 ns/op. Median: 9977 ns/six decisions (about 1663 ns/decision), 6090 B/op, 259 allocs/op. This measures in-process admission only, not lease acquisition or distributed coordination.

Reproduce

go test ./internal/regionadmit -run TestCompareLocalKeepsRegionAlternativesExplicit
go test ./internal/regionadmit -bench BenchmarkDecideSharedRegionAdmission -benchmem -run '^$' -count=5
go test ./internal/nativebench