Lazy lifecycle wrappers for backends without native idle eviction (Phase 1.5)

[jaylfc]

Some backends load their model at process start and have no built-in idle eviction (sd-server, llama.cpp llama-server for the large fallback path, whisper.cpp server). Each needs a small Python proxy that starts the underlying subprocess on first request and stops it after the per-class TTL. Spec: docs/superpowers/specs/2026-04-11-taos-framework-integration-bridge-design.md §Lazy lifecycle wrappers (Phase 1.5).

The sd-server CPU/Vulkan fallback was pinning ~1.6 GB on the Orange Pi with zero recent calls; this wrapper is the correct fix without patching upstream binaries.

Shape:

• Listens on the user-facing port
• Starts the underlying server subprocess on the first inbound request
• Forwards all subsequent requests to the subprocess
• Stops the subprocess after the per-class TTL with no requests; restarts on the next request (pays cold-start once)
• Follows the LLMProxy / QmdClient lifecycle pattern already in the codebase

Acceptance criteria:

• Wrapper implemented for sd-server (CPU/Vulkan SD fallback)
• Wrapper implemented for llama.cpp llama-server (large model fallback path)
• Wrapper implemented for whisper.cpp server
• Each wrapper is ~50 LOC and fully isolated; upstream binaries need no changes
• Wrapper honours the per-class TTL from the scheduler's model registry
• Cold-start latency is logged so the Activity widget can surface it
• Existing smoke tests pass through the wrappers