HydraFusion (policy max): router returns models outside the policy universe, then silently falls back to gpt-5.6-luna

Finding: An external user reports that in the Copilot runtime 1.0.92-4.unstable.r37397886721.gad270aa (bundled SDK 1.0.17-unstable), with VS Code 1.141.0 on a Windows client and a WSL remote, HydraFusion under policy max often fails routing. The runtime rejects the router's chosen model as outside the selected policy universe and then continues on gpt-5.6-luna. Source: https://github.com/github/copilot-cli/issues/5097 Reported evidence: the session log contains session.fusion_route_failed events with reason invalid_response and an error saying the route references a model outside the selected policy universe. In one session over two days the reporter counted 10 failures against 3 successful routes. The rejected models were gpt-6.1-sol (9 times) and claude-sonnet-5.5 (once). The reporter says both models are enabled for the account and listed with policyState enabled in the model list the runtime publishes to VS Code. The UI shows only a brief warning that routing failed and the turn is continuing on gpt-5.6-luna. Additional reported observation: the successful routes under max used only gpt-5.6-luna and gpt-5.6-terra with a Critique pattern. Higher-tier enabled models such as claude-opus-5.5 were never chosen. Those routes report null policyVersion and modelUniverseVersion. The reporter flags this as possibly relevant, but it is a hypothesis, not a confirmed cause. Confirmed cause: unknown. The reporter suspects the server-side router (routeSource capi_plan) uses a different model set than the runtime's max universe, and could not find out how the runtime builds that set. No maintainer response is visible. No workaround, fix commit or released fix is identified in the supplied material. Diagnostic value: if you see the same warning, the session's events.jsonl may contain session.fusion_route_failed entries. Per the report, check the reason and errorMessage fields, and the fallbackModel. The report suggests that a fallback to a lower-tier model means the turn did not run as HydraFusion. Limitations: this is a single external report from one account on an unstable prerelease build. Other policies are untested. The effect of the GitHub Enterprise Cloud data-residency tenant is unknown. Comments were not captured. This is an external report only; KFA did not reproduce it.