Orchestration — multi-model switching & resource policy #16

Planning representation replaced — 6 November 2026. The accepted Measured runtime policy and advanced execution outcome is retained in #62. Reopen one bounded task when selected. This closure does not mean the capability is complete or cancelled. The original scope and evidence below remain preserved; current selected work starts at #52.

Extend the proven chat/Echo/Vision path with measured resource policy and demanding multi-model scenarios.

Entry point: #50 supplies the multimodal acceptance and measurements needed for completion. Policy proposals and fixtures can proceed earlier.

Checklist

• Stress multiple models within one dynamic Soulstone and competing chat/speech/visual runtimes; verify duplicate-load suppression, per-model readiness and whole-Animator lease drain when loading may evict neighbours.
• Define and admit a typed TOML resource policy in ADR 23 using measured headroom and load/switch cost, explicit units/defaults and missing-data behaviour; preserve ownership and hard conflict rules.
• Exercise concurrent requests, priority/FIFO fairness and repeated transitions; measure waiting and switching overhead without silently preempting active work.
• Add a topology-aware contention fixture: one model spans overlapping devices, disjoint work fits, and an extra demand does not. Include stale observations and an observation/admission race; GPUs are not interchangeable slots. Retain an eight-device stress fixture separately from the named physical profile actually tested.
• Test cancellation, partial activation and compensation on a named host; record restored, declined or contained outcomes for Graph/Nexus and Observability / Oculus — correlate one Run across execution #36.

Boundary: Orchestration makes an already selected capability ready. Candidate-selection weights and fallback belong to #17. Current policy does not measure VRAM or switch cost; those are new work.

Connections: Extends #6 and #50; coordinates contracts with #7/#17 without reciprocal completion blockers.

Later: Multi-host placement, durable reservations/fencing, preemption and autonomous recovery, scoped from these results.

References: ADR 23 · ADR 12.

Contribution entry: Contribute a measured runtime scenario or one resource-policy fixture.
Open decision: Choose the first measured policy inputs, units and missing-data rules within ADR 23.
Verification: Offline policy tests plus named-host measurements. Shared contribution and touched-docs rules: #52.