Add OpenShell-backed RSI fleet

[w4ffl35]

What changed

The RSI controller now runs candidate mutations and visible evaluations in NVIDIA OpenShell guests backed by a PostgreSQL job queue and shared S3-compatible artifacts. The queue supports leases, fencing, retries, bounded resource admission, resume validation, and multiple worker processes. The implementation also adds bounded adaptive candidate allocation, fixture-validated curriculum proposals with explicit promotion, adversarial checks, and a QLoRA prototype with paired evaluation.

AirunnerDesktop's Ollama-compatible LLM worker remains the inference route for mutation and can serve the local Qwen GGUF. The supplied GGUF is rejected by the QLoRA trainer; the prototype expects a Hugging Face checkpoint, and no live adapter training has been validated.

Validation

• npm run typecheck
• npm test (all 163 test files passed)
• bash scripts/e2e/run.sh (9/9)
• RSI training Python tests (2/2)
• Local PostgreSQL/RustFS queue load test: 12 simulated workers; 100 jobs; 215.2 jobs/s in the measured drain
• Live OpenShell read-only bind probe: reads succeeded; writes were rejected
• npm pack --dry-run includes the CLI launcher

The local load result does not qualify multi-host throughput. Hidden evaluation and signed evaluator provenance are not implemented. See docs/rsi-progress.md for evidence and limits.