setup: start genie-ai-runtime before memory-heavy services

[carlos4s]

Summary

• raise the default GENIEPOD_AI_RUNTIME_CONTEXT to the Jetson-tested 8192 with --int8-kv
• order genie-ai-runtime.service before Whisper, Home Assistant, and genie-core so runtime claims KV cache before memory-heavy services at boot
• reorder start_all.sh so it starts the configured LLM before Home Assistant and Whisper without stopping those services manually

Fixes #75.

Verification

• cargo fmt --check
• cargo test -p genie-core --test tool_dispatch_test start_all_uses_configured_llm_backend
• cargo test -p genie-core --test tool_dispatch_test genie_ai_runtime_service_preserves_model_page_cache
• bash -n deploy/scripts/start_all.sh
• git diff --check

Jetson context

Issue #75 captured Jetson testing where 8192 loaded with INT8 KV when genie-ai-runtime started before memory-heavy services. This PR preserves that condition through systemd ordering and the lifecycle script, instead of requiring the operator to manually stop the stack first.

[ai-hpc]

Right shape for the right problem. Issue #75's evidence is the smoking gun: same -c 4096 request fits only ~1.7k ctx after the full stack is resident, but fits 4k / 6k / 8k cleanly when genie-ai-runtime loads first. The runtime's -c is clamped by available memory at load time, so startup order materially changes the realized KV-cache capacity.

Three coordinated changes:

1. Before=genie-whisper.service genie-whisper-warmup.service homeassistant.service genie-core.service on the unit. systemd ordering directive that makes the LLM unit's load happen before the other memory-heavy services claim DRAM. Defense in depth with PR #72's After=genie-ai-runtime.service genie-llm.service on genie-core — both sides now agree, and the LLM-vs-whisper / LLM-vs-HA dependencies are also covered. Before= against a unit that isn't installed is a no-op, so this is safe on installs that don't ship homeassistant or whisper.

2. GENIEPOD_AI_RUNTIME_CONTEXT=2048 → 8192. Matches the largest context the issue verified loads cleanly on Orin Nano 8 GB. The env knob stays — operators on smaller Jetsons (Nano 4 GB) can override via systemd drop-in without touching the unit file.

3. start_all.sh reorders so the configured LLM unit + warmup run before homeassistant + whisper in the manual lifecycle path too, mirroring the systemd Before= for operators who use the script directly.

Tests added to tool_dispatch_test.rs lock both invariants:

• start_all_uses_configured_llm_backend extension uses substring-position comparison on the UNITS=(...) array to assert $configured_llm_unit appears before homeassistant.service and genie-whisper.service. Substring-position is fragile (a future rename of the $configured_llm_unit symbol breaks the test), but the assertion message is precise enough that the failure mode points at the right line.
• genie_ai_runtime_service_preserves_model_page_cache extension asserts GENIEPOD_AI_RUNTIME_CONTEXT=8192 and that the new Before= clause is present verbatim.

Worth flagging for a follow-up, not blocking:

• PR #74's GENIE_RUNTIME_MAX_BODY_BYTES = 4 * 1024 body-compaction threshold is now mismatched with the new context. Compaction triggers at 4 KB of body bytes regardless of the runtime's actual capacity, so with 8192-token runtime context (which can comfortably hold ~16-24 KB of typical English chat text), the client will still aggressively compact prompts the runtime could now handle. Same observation I left on #74. Right path is to make the threshold a function of GENIEPOD_AI_RUNTIME_CONTEXT or to probe runtime capacity at connection time. The static 4 KB ceiling is now leaving more performance on the table than it was at 2048-ctx.
• Compatibility with PR #70 (warm page cache) is preserved — Before= only affects boot-time ordering, not systemctl restart of the runtime alone, so the page-cache warmth across restart-of-runtime-only is intact.
• Smaller-Jetson safety: an Orin Nano 4 GB or earlier Jetson without --int8-kv support could fail to allocate 8192 ctx at boot. The unit's Restart=on-failure would loop in that case. Operators on those hosts will need a drop-in: Environment=GENIEPOD_AI_RUNTIME_CONTEXT=2048 (or smaller). Worth a sentence in the next ROADMAP / GETTING_STARTED update if alpha.10 is targeting wider hardware.

All 7 CI checks green on c1cae29 (fmt, clippy, test, aarch64 cross-compile, shellcheck, ruff, --no-default-features). End-user verified on Jetson per your test. Going in.

[ai-hpc]

Merged at 9ab2974e6b0acd3b422576fb57715e71a2086d5b.
Thanks to @carlos4s!