feat(cli): unified `lattice` CLI for users

[ohdearquant]

Context

Lattice has individual binaries (chat_metal, bench_decode_ab, quantize_q4, backfill_qwen3) but no unified user-facing CLI. To use lattice today, a consumer needs to know which binary does what and how to wire features/env vars.

Comparable tools have a single entry point:

• ollama run qwen3.5:0.8b "hello"
• llama-cli -m model.gguf -p "hello"
• mlx_lm.generate --model Qwen/... --prompt "hello"

Goal

A single lattice binary with subcommands covering the common user flows.

Proposed surface (v1)

# Model management
lattice pull qwen3.5-0.8b              # download from HF to ~/.lattice/models
lattice list                            # list local models with sizes
lattice rm qwen3.5-0.8b                 # delete local model

# Inference
lattice chat qwen3.5-0.8b               # interactive REPL
lattice chat qwen3.5-0.8b -p "hello"    # one-shot
lattice complete qwen3.5-0.8b -p "Once" # raw completion (no chat template)
lattice embed bge-small-en -t "hello"   # embedding output

# Quantization
lattice quantize qwen3.5-0.8b --to q4        # safetensors output
lattice quantize qwen3.5-0.8b --to q4-quarot # QuaRot variant

# Server (depends on #92)
lattice serve --model qwen3.5-0.8b --port 8080

# Inspection
lattice info qwen3.5-0.8b                # architecture, params, dtype
lattice bench qwen3.5-0.8b               # decode tok/s, latency p50/p99

Implementation notes

• Use clap derive macros (already in workspace deps)
• Default model dir: ~/.lattice/models (env override: LATTICE_MODEL_CACHE)
• Auto-pull on first use if model not local (with --no-pull to disable)
• Per-subcommand --help with examples
• Single binary in crates/inference/src/bin/lattice.rs (or new crates/cli crate)

Acceptance

• lattice --help shows all subcommands
• lattice chat qwen3.5-0.8b -p "hi" works end-to-end on a fresh machine (downloads, loads, generates)
• Documented in README as the primary entry point
• Existing binaries (chat_metal, bench_decode_ab) become aliases or get deprecated

Priority

P1 — gates real user adoption. Currently lattice is consumable only by Rust code authors who know the crate API.

Related

• feat(serve): long-running daemon for model serving (avoid cold-start) #92 (daemon) — lattice serve is the CLI front-end for it
• feat(serve): OpenAI-compatible chat completions HTTP API #93 (OpenAI API) — server exposes that protocol