feat(compose): optional gpu-coordinator-proxy + ollama services for VRAM contention

[nnnet]

Summary

Adds two optional docker compose services for operators running both LMStudio and Ollama on a single GPU:

• gpu-coordinator-proxy — fronts LMStudio (:1234) and Ollama (:11434) on alternative ports (:1235, :11435) with a shared VRAM lock. Before forwarding to either backend it unloads everything from the other runtime; with GPU_FREE_STRATEGY=wipe-all it can also evict everything from the requested runtime itself for cold-start determinism.
• ollama — official ollama/ollama:latest image with a named volume for model storage, exposing :11434/v1.

The proxy lives in its own repo at https://github.com/nnnet/gpu-coordinator-proxy and is cloned next to MC as ./gpu-coordinator-proxy-src/ (gitignored — the sibling clone is not part of MC's tree). The build.context in the compose file points at that sibling clone.

Why this is useful upstream

Operators running both runtimes on a single dev machine would otherwise have to manually unload models between provider switches. With the proxy, MC and the gateway just point lmstudio.baseUrl=http://host.docker.internal:1235/v1 and ollama.baseUrl=http://host.docker.internal:11435/v1 and the contention is invisible. The 5/6 swap matrix tested cleanly on a 24 GB RTX 5090.

The change is purely additive — omit the two service blocks in docker-compose-openclaw.yml to keep the existing direct-LMStudio / direct-Ollama paths unchanged.

Test plan

1. Clone the proxy: git clone https://github.com/nnnet/gpu-coordinator-proxy gpu-coordinator-proxy-src (next to mission-control/).
2. make up gpu-coordinator-proxy ollama — both services healthy.
3. Pull a model: docker exec mc-ollama ollama pull gpt-oss:20b.
4. Point an MC agent at http://127.0.0.1:1235/v1 (LMStudio) and another at http://127.0.0.1:11435/v1 (Ollama).
5. Dispatch tasks alternating between the two; observe in proxy logs that the other runtime gets evicted on each switch.

Dependencies

• Touches docker-compose-openclaw.yml — same file as feat(openclaw): additive Docker integration with env-driven hardening + doctor cleanup #649 (OpenClaw integration). Reviewers can land in either order; later one rebases.

Provenance

Squashes our fork's commit:

• 3249ee7 feat(compose): add gpu-coordinator-proxy service from sibling repo clone

[0xbrainkid]

Summary

• Blocking: the optional GPU proxy is included as an always-on service with a required local build context, and the compose file also carries a machine-specific OpenClaw state path.

Strengths

• The proxy/ports/eviction-strategy comments are clear and explain the intended LMStudio/Ollama topology well.
• Keeping Ollama model data in a named volume is the right operational default.

Issues

• [BLOCKING] gpu-coordinator-proxy is part of the default compose graph but builds from ./gpu-coordinator-proxy-src, which is intentionally gitignored and absent on a normal clone. docker compose up / make up will fail unless the operator manually clones the sibling repo or edits the compose file, even though the comments call the proxy optional. Please put this service behind a profile (for example profiles: ["gpu-proxy"]) or split it into an optional override compose file so the default OpenClaw stack starts without the external repo.
• [BLOCKING] Same portability issue as the OpenClaw integration compose: OPENCLAW_STATE_DIR, OPENCLAW_PLUGIN_STAGE_DIR, OPENCLAW_CONFIG_PATH, and the state bind mount hard-code /mnt/9/gt/rig_PlatformsAI/mayor/rig/beads/discovered/mission-control/.openclaw-data. That path will be wrong for other clones/hosts. Please make it configurable/generated locally rather than committed.

Questions

• Should Ollama also be profile-gated, or is it intended to be part of the default OpenClaw stack even for cloud-only users?

Verification

• gh pr checks 650: no checks reported.
• docker compose -f docker-compose-openclaw.yml config --quiet: passed syntax/config rendering, but does not validate that the optional build context exists or that the committed absolute path is portable.

[0xNyk]

Thanks — holding this one until #649 lands. They both modify docker-compose-openclaw.yml (this PR adds the gpu-coordinator-proxy and ollama services to the file that #649 introduces), so the merge order matters and the diff is currently hard to evaluate in isolation.

Once #649 lands, please rebase this onto the new main so the diff shows just the GPU-coordinator delta (the actual ~600 lines you're adding), not the full file. That'll make this trivially reviewable.

One question while you're rebasing:

The proxy service description mentions GPU_FREE_STRATEGY=wipe-all for cold-start determinism — that strategy evicts everything from the requested runtime as well, which on shared LMStudio/Ollama setups would kill any other concurrent workload. Could you add a note in the compose comment that operators on multi-user GPU hosts should leave it on the default wipe-other strategy? Just so the trade-off is visible at the spot where someone enables the flag.

Tagging as blocked on #649 for now. The technical content here looks sensible — VRAM contention is a real problem and a coordinator proxy is the right shape for it.