Residency unification, disk GC, worker-side VRAM juggling (#369, #370, #371)

[PaulFidika]

Closes tracker issues #369, #370, #371 (agents/completed.md).

#369 — residency registry owns the executor's pipelines

• Per-slot allocator-delta measurement in _injection_kwargs: each worker-built pipeline is registered under its ref with its own measured vram_bytes AND the object (demote/promote actually move memory). Tenant-loaded refs (path-annotated setup slots) split the residual delta — multi-model endpoints no longer report vram_bytes=0.
• Model loads serialize under Executor._load_lock — concurrent setups can no longer cross-contaminate allocator deltas or place_pipeline's free-VRAM read.
• demote()/promote() only perform transitions they can actually execute (movable object + host-RAM headroom); otherwise the executor vacates the owning record and books every held ref back to disk — registry state and instance state cannot diverge.
• Free-VRAM probes (residency + StateDelta) sum across ALL CUDA devices.
• Bug found en route: residency events raised from worker threads were silently dropped when no loop had been captured (ModelStore.bind_loop).

#370 — keep-set disk retention (disk GC exists now)

• Pre-download headroom gate in ModelStore.ensure_local (snapshot sizes are known up front): GC first, then ModelEvent{FAILED, insufficient_disk} + typed InsufficientDiskError (fails the job RETRYABLE without self-disabling the function).
• LRU disk GC over Residency DISK-tier refs honoring keep + in-use pins + a 300s grace window; keep-pressure escape hatch evicts LRU keep refs too, still emitting EVICTED.
• Deletion units in models/disk_gc.py: CAS snapshot dir + hardlink-count orphan-blob sweep; HF models--* repo dir.
• Persisted ref-index.json + boot-time rescan wired into lifecycle.startup() so Hello.models matches disk truth across restarts.

#371 — worker-side VRAM juggling

• ensure_setup runs make_room before loading (estimate: per-ref vram_hint, else declared resources.vram_gb); idle LRU pipelines demote to the warm RAM tier instead of the new load degrading down the offload ladder. Non-movable victims vacate via record teardown. Offload ladder remains the fallback.
• Hub UNLOAD demotes instead of destroying (instance stays ready); the next RunJob/LOAD promotes RAM→VRAM in seconds.

Tests

uv run --extra dev pytest -q: 193 passed, 1 skipped (was 182+1s on master).
New: tests/test_executor_residency.py (6: per-ref measured bytes, residual attribution, concurrent-setup non-contamination, UNLOAD→RAM→promote round trip, tenant-loaded teardown, alternating demote/promote swap with IN_VRAM⇄IN_RAM events), tests/test_disk_gc.py (5: LRU non-keep eviction with real file deletion, keep escape hatch, fail-fast insufficient_disk, in-use pin, restart rescan).