[joelteply]

Serves the residents' health on every kind of machine (Joel, 2026-11-10: "It depends on where it's resident... We just can't assume any specific machine"). Two commits.

1. An embedding is charged where it resides, and the lane's floor is what it embeds

• LlamaCppBackend::context_pool decides where a context lives:
  • Device memory when llama offloads to a GPU (llama::supports_gpu_offload, new) and the backend offloads any layer.
  • Otherwise host memory, whatever the machine.
• inference/embed_lane_memory.rs owns the lane's memory:
  • Each lease is charged to that pool.
  • The standing floor is re-stated at each lease, at the largest context the lane has embedded. The resolver's probe and calibration passes set it before any turn.
• Deleted: EMBED_LANE_VRAM_BYTES (a fixed 1,792 MiB VRAM bound) and the resolver's reservation.
• llamacpp_adapter.rs shrinks from 938 to 848 lines.

2. A turn whose grant is being reclaimed ends instead of waiting on the reclaim it blocks (fixes #5906)

• A queued mind waits her turn; an embedding leases what its own context allocates #5906 made a Reclaiming grant wait inside her turn.
• The reclaim takes her turn lock (MindTurn::reclaim), so the turn waited on itself.
• On the M5 (d0abbdd), Aris parked to the 1,500 s tick deadline five times and made 3 acts in 132 minutes.
• Now Reclaiming refuses again. Queued and an engine relaunching still wait.
• Each wait probes its reason (delib.window.waiting).

Measured

• Local tests: continuum-engine 1,527 passed on the live stack (serving-down window, 13:30Z).
• Live: deployed on the M5 at 13:32:50Z (d613517). The 15-minute read, including Aris's acts, follows in a comment.

Not here

• Residence tiers (device, host, flash: mmapped weights, KV pages on disk, streamed experts) as one accounting, and the VRAM board double count on shared memory. Proposed to Codex; it feeds the serving plan.
• The per-token rate from the fork: card 9bddda93.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LoTjvf5j3Ez13g6k8mRkFo

[joelteply]

Third commit: a mind's children run below the engine that serves her (continuum-host process_class, applied in the kernel's command_for). On Unix that's the OS's lowest ordinary priority; on Windows, the idle class. Measured before: a resident's full-workspace test run at the engine's priority pushed load to 36, swapped 4.5 GB, and dropped decode to 3.4 tokens/s (56 at rest). Tests: continuum-host 17 passed (including a child reporting a lower priority than its parent) and continuum-kernel 1,403 passed. Live on the M5 as caf627c. This is the interim contention fix; the full design is card 13c81ced.

[joelteply]

Included in the live stack, #5917 (one line of work: no duplicate copies). #5913's board change stays reverted there until a serving-plan replay proves it.