v0.2: Primary slot served by Lemonade end-to-end (tracer bullet)

[thinmintdev]

What to build

Single-model end-to-end vertical slice. A fresh hal0 box installs Lemonade, brings up lemond under systemd, and serves a chat completion via Lemonade for the primary slot when HAL0_BACKEND=lemonade flag is set. v0.1.x toolbox path remains default + working.

Touches every layer end-to-end:

• install.sh — download AMD's embeddable tarball, sha256-verify, extract to /opt/lemonade, apt-install unzip + libxrt-npu2, run lemonade backends install llamacpp:rocm at first boot
• systemd — write /etc/systemd/system/lemond.service running lemond /opt/lemonade --port 9100 with hardening directives (NoNewPrivileges, ProtectSystem=strict, ProtectHome, PrivateTmp, RestrictAddressFamilies)
• manifest.json schema v2 — lemonade: { tarball_url, sha256, version }
• src/hal0/lemonade/client.py — minimal HTTP client wrapping /v1/load, /v1/unload, /v1/health, /v1/chat/completions, /live. /v1/load body shape: {"model_name": "..."} (only required field, per research)
• src/hal0/providers/lemonade.py — LemonadeProvider for the primary slot only; delegates to LemonadeClient
• HAL0_BACKEND=lemonade flag — env var or CLI flag plumbing; old LlamaServerProvider is default; new path only when flag set

Demoable: SSH to test box → set flag → /v1/chat/completions returns a non-empty response from a Lemonade-served primary slot.

Acceptance criteria

• Fresh install with HAL0_BACKEND=lemonade brings up lemond.service healthy via /live
• LemonadeClient.load("qwen3.5-0.8b") returns success; /v1/health shows the model loaded
• /v1/chat/completions proxied through hal0 returns a valid streaming response
• Default path (no flag) still works — v0.1.x toolbox primary slot unchanged
• manifest.json v2 schema validates
• Integration test asserts the full path on CI (or marked needs-hw if Strix Halo only)
• unzip + libxrt-npu2 prereqs handled in install.sh idempotently

Blocked by

• v0.2: Land Lemonade ADRs + migration plan + research docs #139

[thinmintdev]

In-flight via PR #137

PR #137 (open) implements the LemonadeClient portion of this tracer bullet:

• Control-plane HTTP wrapper (src/hal0/lemonade/client.py)
• Conforms to ADR-0006 §3 post-research shape (HTTP-only, no CLI fallback, /v1/load body is {"model_name": "..."})

Remaining scope to land THIS issue:

• LemonadeProvider for primary slot only (consumes feat(lemonade): client skeleton — control-plane HTTP wrapper (ADR-0006 PR-2/16) #137's client)
• install.sh — tarball fetch + sha256 + apt prereqs + lemonade backends install llamacpp:rocm at first boot
• /etc/systemd/system/lemond.service (hardened bare process)
• manifest.json schema v2 lemonade: { tarball_url, sha256, version }
• HAL0_BACKEND=lemonade flag plumbing
• Integration test asserting end-to-end chat completion

Cadence note: the parallel session originally framed this as a 16-PR plan; that's been re-sliced into the 10 vertical-slice issues #139-#148. PR #137 reads "PR-2/16" in its description — that naming is from the older plan but the underlying code aligns with post-research ADR-0006.

[thinmintdev]

Client piece landed — PR #137 (b52b963) merged

src/hal0/lemonade/{client.py,errors.py,__init__.py} + tests on main. Specialised LemonadeLoadError for the evict-all blast radius (ADR-0007). /live returns bool/never raises. No auto-retry on /v1/load.

Remaining scope for this issue:

• LemonadeProvider for primary slot (src/hal0/providers/lemonade.py) — consumes LemonadeClient
• install.sh — tarball fetch + sha256 + apt deps + first-boot lemonade backends install llamacpp:rocm
• /etc/systemd/system/lemond.service — hardened bare process
• manifest.json schema v2 — lemonade: { tarball_url, sha256, version }
• HAL0_BACKEND=lemonade flag plumbing
• Integration test: chat completion through Lemonade end-to-end