Rebuild eval architecture: 3-oracle groundedness, composite scoring, RAG, hotkey-only validators

[kyron1112567]

• feat: enhance leaderboard with submission metadata including agent details and artifact SHA
• feat(orchestrator): chat-stream passthrough; consumer-api becomes thin SSE facade
• feat(eval): multi-turn fixtures with live + scripted replay
• feat(dashboard): queued submissions page
• feat(cost): per-task LLM cost via owner-api ledger injection
• fix(validator): burn 1.0 - total_assigned so EIREL_FAMILY_WEIGHTS controls real burn
• chore(fee_verifier): boot-time audit log + per-call DEBUG
• feat(monitoring): add kube-state-metrics and node-exporter configurations
• refactor: remove honeytoken and trace gate penalty features from the codebase
• feat(schema): add EvalFeedback + OrchestratorToolCallLog tables; per-dimension TaskMinerResult fields
• feat(shared): add safety pipeline + multi-metric scoring + tool pricing table
• feat(tools): server-attested tool-call ledger; url-fetch + mcp-relay services; sandbox session persistence
• feat(rag): rag-tool-service + per-run corpus indexer + miner-pod EIREL_RAG_URL injection
• feat(validator): 3-oracle layer (OpenAI+Gemini+Grok) + Chutes reconciler + oracle cache
• feat(validator): composite scoring with hard gates; pairwise from cached oracle answer; hotkey-signed ledger fetch
• feat(orchestrator): graph-based product runtime; serving picker; promotions; mcp dispatcher
• feat(owner-api): hotkey-signed eval feedback + ledger; tier-1 leaderboard fields; closed-run public submission viewer; auth tightening
• chore: drop legacy x-tool-service, OpenAI baseline client, calibration manager, datasets router
• refactor(shared): drop dead judge_config plumbing from Settings + bundle; tolerate legacy oracle_source values
• chore(deploy): consolidate to EIREL_EVAL_JUDGE_*; drop validator EIREL_INTERNAL_SERVICE_TOKEN + EvalFeedback proxy envs; wire rag-tool-service
• chore: pyproject script entries (rag-tool-service, url-fetch); ensure dataset dir tracked
• docs: README + miner-guide + validator-guide for the new eval architecture
• test: oracles + reconciler + composite + ledger + feedback + product orchestrator + safety coverage