[codex] Support local hardware profile overlays

[TheEditor]

Summary

Adds a local hardware profile overlay layer. Oathweaver can now load host-specific profile definitions from SourceCode/configs/hardware_profiles.local.json and merge them over the shared hardware profile config at startup.

This preserves the project’s shared runtime defaults while giving operators a documented path for local capacity tuning that does not require editing tracked config files.

Why

The active hardware profile layer introduced in PR #2 makes runtime capacity explicit through shared config. That is useful for portable defaults, but individual deployments still need a place for machine-specific profile values, local defaults, and temporary tuning.

This PR adds that local-only layer so future users can describe deployment-specific capacity without carrying dirty changes in SourceCode/configs/hardware_profiles.json or proposing machine-specific profiles upstream.

Changes

• Adds optional local profile loading from SourceCode/configs/hardware_profiles.local.json.
• Merges local profile definitions over the shared profile payload when present.
• Allows the local overlay to override default_profile for a deployment.
• Ignores SourceCode/configs/*.local.json so local profile overlays stay out of source control.
• Documents the local hardware profile workflow in the README.
• Updates the Linux web launcher so --no-restart-ollama reuses the configured Ollama endpoint instead of starting another ollama serve process.
• Adds tests for adding local profile definitions and overriding the active default through the local overlay.

Example Profile Shape

{
  "default_profile": "local_8gb_vram_48gb_ram_cuda",
  "profiles": {
    "local_8gb_vram_48gb_ram_cuda": {
      "name": "local_8gb_vram_48gb_ram_cuda",
      "display_name": "Local: 48GB RAM / 8GB CUDA",
      "hardware": {
        "system_ram_gb": 48,
        "gpu_backend": "cuda",
        "gpu_vram_gb": 8,
        "unified_memory": false
      },
      "scheduler": {
        "max_context_tokens": 12288,
        "warning_context_tokens": 8192,
        "max_stage_context_tokens": 2400,
        "max_parallel_models": 1,
        "max_active_model_calls": 1,
        "on_deck_depth": 1,
        "warm_depth": 0,
        "allow_neural_prefetch": false
      }
    }
  }
}

Validation

• PYTHONPATH=SourceCode python3 -m unittest tests.test_hardware_profiles tests.test_inference_router_compat tests.test_model_routing_tier_resolution tests.test_ollama_wait_for_available
• PYTHONPATH=SourceCode .venv/bin/python -m unittest tests.test_hardware_profiles
• python3 -m json.tool SourceCode/configs/hardware_profiles.json >/dev/null
• python3 -m json.tool SourceCode/configs/hardware_profiles.local.json >/dev/null
• bash -n start_oathweaver_web.sh
• git diff --check