Make hardware profile an explicit runtime policy

[TheEditor]

Summary

Adds a first-class active hardware/runtime profile layer. Oathweaver can now resolve a named hardware profile at startup, initialize scheduler budgets from that profile, and pass normalized capacity policy into inference-router diagnostics.

This makes the project’s local runtime assumptions explicit without changing model routing intent or adding machine-specific profiles.

Why

Oathweaver already has hardware-aware pieces: scheduler budgets, benchmark profiles, model-routing diagnostics, premium model locks, and conservative fit estimation. Those pieces were not yet connected through one durable active runtime policy.

This PR makes the target-machine policy explicit so future users can describe local capacity in config instead of relying on implicit workstation assumptions.

Changes

• Adds structured hardware profile config in SourceCode/configs/hardware_profiles.json.
• Adds active profile resolution via OATHWEAVER_HARDWARE_PROFILE.
• Converts resolved profiles into HardwareBudgetProfile for scheduler/resource-budget use.
• Converts resolved profiles into router policy for estimate_fit() and validate_config().
• Includes active hardware profile details in pipeline and replay diagnostics.
• Allows kernel commands to apply structured hardware profiles by name.
• Keeps enforcement advisory-first: profile mismatches warn rather than hard-block startup or model calls.

Example Profile Shape

{
  "name": "local_cuda_workstation",
  "display_name": "Local CUDA Workstation",
  "description": "Example profile for a local CUDA machine. Values are illustrative and should be adjusted by the operator.",
  "hardware": {
    "system_ram_gb": 32,
    "gpu_backend": "cuda",
    "gpu_vram_gb": 12,
    "unified_memory": false
  },
  "scheduler": {
    "max_context_tokens": 8192,
    "warning_context_tokens": 6144,
    "max_stage_context_tokens": 2200,
    "max_parallel_models": 1,
    "max_active_model_calls": 1,
    "on_deck_depth": 1,
    "warm_depth": 0,
    "allow_neural_prefetch": false
  },
  "inference": {
    "preferred_backends": ["llama.cpp", "ollama"],
    "default_keep_alive": "5m",
    "heavy_keep_alive": "0",
    "release_heavy_after_call": true,
    "max_loaded_models": 1
  },
  "model_policy": {
    "normal_max_b": 9,
    "heavy_max_b": 14,
    "premium_min_b": 24,
    "allow_premium": false,
    "premium_requires_manual": true,
    "allow_14b_with_warning": true,
    "reject_heavier_fallbacks": false
  },
  "lane_caps": {
    "chat_layer": {
      "max_context_tokens": 6144,
      "max_parallel_agents": 1
    }
  },
  "validation": {
    "startup_mode": "warn",
    "strict_mode_available": true,
    "warn_on_missing_models": true,
    "warn_on_unreachable_backends": true,
    "warn_on_context_over_cap": true,
    "warn_on_parallelism_over_cap": true,
    "warn_on_premium_auto_escalation": true
  }
}

Validation

• python3 -m unittest tests.test_hardware_profiles tests.test_inference_router_compat tests.test_phase12_cag_bench_integration
• python3 -m unittest discover tests
• python3 -m json.tool SourceCode/configs/hardware_profiles.json >/dev/null
• git diff --check