Reuses the existing provision-raw-gpu-droplet.sh + model-boss coordinator;
verified backend tsc clean, 298 tests (17 new), web build green.
- src/gpu/: DO droplet lifecycle (fetch-based DO v2 client, provision /
idle-teardown / status), model-boss client with a connectivity circuit
breaker, gpu_droplets state (migration 0012) reconciled against DO on
startup. GET /prospector/gpu/status + POST provision/teardown.
- GpuEnrichedClassifyService: async 22-atom classify + Quinn-voice draft
via model-boss, returns null on any failure. Wired into the task runner
(advanceClassify/advanceDraft) and /classify (when handle present) with
silent fallback to the fast classifier + pastebin. The synchronous
inbound gate is untouched; GpuModule is a DAG leaf.
- Boot-clean without DO_API_TOKEN / MODEL_BOSS_URL (enrich disabled,
provision throws a clear error). All secrets via ConfigService.
- web Hosts view (designs/hosts-do-gpu.html): droplet card + model-boss
reachability + provision/teardown + mesh/fleet sidebar.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
b9bc5964cf