feat(serve): long-running daemon for model serving (avoid cold-start)

[ohdearquant]

Context

Loading Qwen3.5-0.8B weights takes ~2-5s (memory map + Metal buffer upload). For batch workloads (embedding pipelines, RAG, evaluation) this is acceptable. For interactive use it's painful — every CLI invocation pays the load cost.

Ollama solves this with a daemon that holds models in memory and serves multiple requests. Lattice needs the equivalent.

Goal

A long-running lattice daemon process that:

1. Loads models on demand (first request → load, subsequent → reuse)
2. Supports multiple models in memory simultaneously (LRU eviction when GPU/RAM constrained)
3. Streams tokens via stdout, HTTP, or Unix socket
4. Survives multiple client invocations without re-loading

Proposed design

Process model

• lattice serve [--port 8080] [--socket /tmp/lattice.sock] starts the daemon
• Daemon runs in foreground (Ocean's preference — no systemd entanglement); user can & or nohup it
• Lock file at ~/.lattice/serve.pid prevents double-start

Storage

• Loaded models held in a Arc<RwLock<HashMap<ModelId, LoadedModel>>>
• LRU eviction when total VRAM > configured limit (default: 80% of detected Metal allocator size)
• Per-model KV cache pool (issue feat(serve): OpenAI-compatible chat completions HTTP API #93)

Protocol

v1: simple line-protocol over Unix socket (newline-delimited JSON requests + SSE-style token stream responses). Easy to debug with nc.
v2: HTTP + OpenAI API (issue #92).

Concurrency

• Single inference at a time per model (Metal isn't trivially shareable across threads)
• Multiple model serves can run concurrently if they fit
• Request queue with backpressure (429 if queue full)

Acceptance

• lattice serve + lattice chat ... --daemon round-trip works
• Sub-second response start for cached models (vs current ~2-5s cold load)
• Survives 100 sequential requests without leaking
• Handles SIGTERM gracefully (finish in-flight request, then exit)

Priority

P1 — blocks daily usage. Tied to #91 (CLI) and #93 (OpenAI API).

Related

• feat(tune): Adam/AdamW optimizer + LoRA gradient computation #90 — CLI entry point (lattice serve)
• feat(serve): long-running daemon for model serving (avoid cold-start) #92 — OpenAI-compatible HTTP API on top
• bench: cross-framework performance comparison suite #84 — cross-framework bench; daemon makes lattice fair to compare against ollama