docs: record replay acceptance and bounded factory operating review

docs/assessments/2026-09-09-helixforge-replay-and-operating-contract.md

@@ -0,0 +1,58 @@
+# Factory continuation: replay acceptance and operating contract
+
+2026-09-09. Programme: HFACT-WP-0001 in prj-helixforge-factory.
+[Machine-readable return](2026-09-09-helixforge-replay-and-operating-contract.json).
+
+The Secrets Engine consumer now implements FLEX-DEC-2026-012. It binds an allow
+to the exact submitted request, including its approval claim, compares the two
+evaluator-origin approval digests, rechecks freshness after Check, and consumes
+the evaluated request digest with its decision id before protected effects.
+334 regression tests and 13 real local KeyCape/Approval Engine/Flex Auth checks
+passed. Registry override, changed submission, wrong action, real dual-control
+policy, consumed approvals and CAS retry/conflict are covered. The former PDP
+double limitation is resolved for component conformance; no live native delivery
+or governed factory run is claimed. The historical live fixture remains intact
+and refuses because it lacks the new submitted binding.
+
+Implementation: Secrets Engine `ee4e901`; published final source and exact
+Railiance primary reconciliation are in the machine-readable return.
+SECRETS-WP-0008-T02 is wait for current deployed PDP/approval-path adoption.
+SECRETS-WP-0009-T03 keeps native delivery; RPF-WP-0035-T06, AUDIT-WP-0009-T09
+and APPROVAL-WP-0002 keep client-side/audit/service admission. Completed verifier
+CCRs were neither reopened nor treated as broader read authority.
+
+The [bounded operating packet](../../../prj-helixforge-factory/operations/bounded-operating-contract.md)
+brings G0 preparation forward. Its proposed identity/profile/host, accountable
+owners, one-commit v1 path grant, useful demand and independent oracle are
+explicit. The actual rein parser accepts the grant and rejects publish=true.
+The proposed demand extends reuse-surface with typed hosted-discovery refusals
+for factory intake and vergabe-teilnahme's delivery checks. Product selection
+remains reuse-surface first, vergabe-teilnahme as customer service/UI, Railiance
+Fabric for later placement utility. The packet is a blocked draft, not an
+execution or spending grant.
+
+Source review identified two precise returns before admission:
+
+- HFACT-WP-0001-T01: the Claude adapter's before/after token bookkeeping does
+  not establish a hard monetary cap. Prove an actual provider/runtime spend
+  bound, including retries and persistent daily/total reservations, before
+  paid execution and final operating acceptance.
+- HFACT-WP-0001-T05: the outer worker already checks grants and writes durable
+  metrics. The selected sandbox mirrors the checkout; its commit must be
+  preserved and imported under the original baseline/lease/transaction before
+  teardown. Returning a commit identifier cannot satisfy host acceptance. Keep
+  existing checks/outbox and prove this actual artifact join with the owners.
+
+These findings refine the original integration plan. The next increment should
+address those two mechanisms alongside existing credential/audit/service
+admission, then accept the exact operating/profile/placement tuple and execute
+one real queue run. Broad historical cleanup and extra factory frameworks do
+not become new prerequisites.
+
+Efficiency improved by removing an unsatisfiable consumer contract, testing
+three actual components together, and identifying runtime gaps before an
+attended model/deployment attempt. No elapsed-time savings or production
+throughput are inferred. The factory ledger still contains zero governed
+attempts; all G0-G5 acceptance gates remain open and the fourteen-day observation
+window has not begun. Source-backed task fields are verified after publication;
+any repaired projection omissions remain an unresolved Repo Manager defect.