deploy: opt-in genie-ai-runtime backend in setup + systemd (#27)

[enjoyandlove]

Summary

Closes #27

• Wire [services.llm].backend (llm: select backend from config #43, runtime: resolve llm service from config #40, core: expose configured llm backend #39) into the deploy layer so opt-in installs validate the right binary and enable the right systemd unit instead of silently falling back to llama-server / genie-llm.service.
• Add deploy/systemd/genie-ai-runtime.service for jllm-server, mirroring genie-llm.service flag-for-flag so the OpenAI-compatible client in genie-core works against either backend unchanged.
• Document the selector, accepted values, and migration steps in doc/configuration.md; list the new unit in doc/deployment-and-ops.md.

Changes

• deploy/setup-jetson.sh
  • Parse [services.llm].backend and [services.llm].systemd_unit from geniepod.toml with a tolerant awk reader (strips quotes, trailing comments, whitespace; accepts hyphenated aliases).
  • Step 5 now branches on the configured backend: checks llama-server for llama_cpp (default), jllm-server for genie_ai_runtime, and warns on unrecognized values without aborting.
  • The systemd enable loop and the final "Start services" hint resolve the LLM unit from config instead of hardcoding genie-llm.
• deploy/systemd/genie-ai-runtime.service — new unit for /opt/geniepod/bin/jllm-server. Same flag surface as genie-llm.service, gated by ConditionPathExists so it no-ops when the binary isn't installed.
• doc/configuration.md — [services.llm].backend table, hyphenated-alias note, and the three-step flip checklist (set systemd_unit, install binary, re-run setup).
• doc/deployment-and-ops.md — add genie-ai-runtime.service to the systemd unit inventory.

Test plan

• On a config with no [services.llm].backend, setup-jetson.sh reports Checking llama.cpp backend (configured: llama_cpp) and enables genie-llm.service (regression — default path unchanged).
• Set [services.llm].backend = "genie_ai_runtime" and systemd_unit = "genie-ai-runtime.service"; re-run setup — script reports the genie-ai-runtime check, enables genie-ai-runtime.service, and the final hint prints sudo systemctl start genie-ai-runtime.
• With backend = "genie_ai_runtime" but jllm-server missing, setup prints the build/install instructions and the revert hint, and continues (does not abort).
• Set backend = "bogus" — setup emits the WARN line and skips the binary check without failing.
• Hyphenated alias "genie-ai-runtime" is accepted and behaves identically to the underscored form.
• systemctl cat genie-ai-runtime.service matches the new unit; ExecStartPre cache-drop and SupplementaryGroups=video render are present.

[ai-hpc]

Closing as superseded by the work that actually landed on main, not because anything was wrong with this PR — the design here is solid and matches the issue's opt-in plan exactly. Timing just played out differently.

Between this PR opening on 2026-05-17 and now, the runtime swap landed in a different shape via several smaller PRs:

• [services.llm].backend end-to-end wiring — llm: introduce backend facade #35, voice: report configured llm backend #38, core: expose configured llm backend #39, runtime: resolve llm service from config #40, llm: select backend from config #43. The runtime-side LLM facade + backend selector this PR was going to consume in the deploy layer.
• deploy/setup-jetson.sh backend-aware enable loop — Fix: Default to Genie-ai-runtime #55 (default flip to genie_ai_runtime + Qwen3-4B), with section-aware awk reading of [services.llm].backend matching your design. Hyphenated-alias handling kept.
• genie-ai-runtime.service unit — feat(deploy): genie-ai-runtime v1.0.0 build+install pipeline (issue #54) #56 (runtime build/install pipeline, with --runtime genie-ai-runtime flag) and setup: start genie-ai-runtime before memory-heavy services #76 (boot-ordering Before= + --int8-kv -c 8192 for the Jetson-tested context).
• Auto-fallback when jetson-llm-server is missing — fix(setup-jetson): auto-fallback to llama.cpp when jetson-llm-server is missing (#60) #63 (closes setup-jetson.sh enables genie-ai-runtime.service without verifying jetson-llm-server is installed #60, an issue your PR predicted with the "if binary missing, continue with revert hint" check).
• Page-cache preservation across restart — setup: preserve warm genie-ai-runtime model cache #70 (closes Investigate slow Qwen3 model loading and warm weight behavior #69, dropped the ExecStartPre cache drop from your unit shape).

Net effect: [services.llm].backend = "genie_ai_runtime" is the v1.0.0-alpha.9 default, genie-ai-runtime.service is on disk and enabled, setup-jetson.sh reads the configured backend and enables the right unit, and the auto-fallback path catches the "binary missing on fresh install" case. The opt-in phase from #27 got skipped because genie-ai-runtime v1.0.0 dropped before the alpha.8 GenieClaw cut and we went straight to default-with-fallback.

Closing #27 separately with a roll-up of where each acceptance criterion ended up. Thanks for the careful design on this one — the awk parser pattern in particular got picked up and used in the later PRs.