Simultaneous HTTP workload release admission
This bounded model adds the HTTP handler interleaving that the store-level concurrent-admission model abstracts away. Two handlers can both miss the initial stable-key lookup and read mutable qualification, policy, route, and source evidence. The model then treats AcceptPrivateWorkloadRelease as one atomic store boundary: one caller creates the durable acceptance and the other receives that acceptance as a replay. It checks that there is at most one fresh operation and that both successful callers identify that same operation.
Regenerate
From the repository root:
~/.agents/skills/statechart/scripts/statechart \
docs/v3/statecharts/workload-release-http-concurrency/AdmissionHTTP.tla \
--config docs/v3/statecharts/workload-release-http-concurrency/AdmissionHTTP.cfg \
--view docs/v3/statecharts/workload-release-http-concurrency/view.ts \
--out /tmp/norn-workload-release-http-concurrencyThe checked config reaches 19 states and 28 transitions. TypeOK, AtMostOneFreshAcceptance, FreshAcceptanceHasStableSelection, EveryReplayReturnsAcceptedSelection, EveryFreshResultHasDurableAcceptance, BothSuccessfulRequestsConverge, and deadlock checking pass. The chart hides four redundant actions and six transitions that stay within one displayed state.
Runtime correspondence and limits
TestEtcdWorkloadReleaseAdmissionConcurrentSameKeyConverges starts two real HTTP handler requests together and holds both at the initial replay miss. Both then traverse the trusted-evidence path and call the acceptance boundary; the test checks one fresh response, one replay response, and the same operation ID in both. It uses a deterministic fake store whose acceptance boundary models one atomic winner. Separately, TestV3PrivateWorkloadReleaseAcceptanceIsAtomicReplaySafeAndKeyRecoverable uses two independent etcd clients to prove the real store boundary returns one fresh acceptance and one replay under the same race.
Run the HTTP interleaving test with:
cd v2/api && go test -race . \
-run 'TestEtcdWorkloadReleaseAdmissionConcurrentSameKeyConverges' -count=1These two proofs cover adjacent boundaries, not one end-to-end HTTP-to-etcd test. The model assumes atomic store acceptance; the HTTP test uses a fake at that boundary, and the etcd test does not pass through the HTTP handler. It also abstracts qualification freshness and concurrent reads, process restart, private-input rotation/deletion, worker recovery, Nomad/ingress drivers, and live workload behavior. Those gates remain open.
An initial disposable-etcd HTTP harness was discarded because it tried to seed an already-succeeded qualification deployment directly. The real store requires that source to have completed signed deployment and ingress-route admission, so direct etcd writes would make the concurrency result misleading. The replacement needs to create that source through the real route-completion path; only the Nomad lineage observer may be substituted. See the dated fixture-boundary note in the M5 workload-release protocol.