Tests for M6 provisioning, stable identity bindings, receiver authorization/marking, restricted gateway configuration and OAuth sessions, with retained real-model fixture drivers.
The added real Git regression stages a new payload, simulates a swallowed submit and stale queue label, then checks that the real CLI retains the unmarked record. The positive control clears the composer and proves the durable mark. The old unit assertion accepting a label alone was an incorrect oracle: it now rejects that frame and separately accepts the label with an empty composer. Existing payload/redraw protection tests remain intact.
Current tips: product 586bd10eafc42b32f4b80de439060c92e447ab96; tests 3c85b0a0b0fc0d773d1b869b06a8c728392fb0e0.
Correction validation on b1: 248 Python tests passed in 56.128s, zero skips; real Git boundary failed before the fix, then passed stale-label refusal with zero marks/one pending record and positive empty-composer acceptance with its durable mark. Source at /tmp/m6-builder-revision matches these committed tips.
Earlier M6 evidence: both real Codex queue transport gates passed A→B→A, offline recovery, failure/refusal preservation and full lifecycle restart. Helper 2 + messenger 14 + spike 1 Go tests/builds and four Nix live suites passed previously; these and the model sessions were not rerun for this tmux-only correction. Messenger a212901, Habitat/SAP 242714e, running/profile k0nxcil6 remain unchanged. Independent adversary review is still required.
Product: #3. This PR targets that product branch.