Centralize config via Pydantic/dataclass and env overrides

[vecnode]

Replace scattered SimpleNamespace mutations with one validated settings object; document keys in README. Reduces bugs from typos and drift.

[vecnode]

Quick status + next steps for this config centralization work:

What is in place already

• The codebase has started moving toward more structured configuration for runs (e.g., via templates and clearer entrypoints), which reduces some of the previous ad-hoc SimpleNamespace usage.
• Recent issues and refactors are already assuming a single, coherent config object that can be passed through the pipeline, which aligns well with this issue’s goal.

What remains to fully resolve this issue

• Implement a single Settings object (Pydantic model or dataclass) that holds all top-level configuration for a run, including model, sampler, device, IO paths, and any RTX/profile presets.
• Replace remaining scattered SimpleNamespace or dict mutations with reads/writes on this Settings object, ideally making it mostly immutable once a run starts.
• Add environment-variable overrides in a thin layer that populates or amends Settings before the run begins (e.g., DISCO_DEVICE, DISCO_PROFILE, etc.).
• Document all keys on this Settings object in the README or a short config reference, so that templates, API, and CLI usage all speak the same “config language”.

Once this is done, it will also make it much easier to implement things like RTX fast profiles and template-driven runs without duplicating config logic across entrypoints.

[vecnode]

Small, reviewable first step toward centralized config

Goal
Introduce a tiny RunConfig dataclass in a dedicated module, and use it in one place (no behavior change), as a first step toward centralized config.

1) Add a config module
Create src/disco/config.py (path adjusted to the layout if needed) with something like:

from dataclasses import dataclass
from pathlib import Path
from typing import Optional

@dataclass
class RunConfig:
    """Minimal, first-step configuration object for a diffusion run.

    This groups a few core settings that were previously passed around
    as separate arguments. It is intentionally small so it can be
    extended toward a full Settings object for issue #2.
    """
    output_dir: Path
    device: str = "cuda"
    seed: Optional[int] = None

Pick 2–3 fields that are already clearly present in the existing entrypoint (e.g., output_dir, device, seed) and match names/types to current usage so there is no behavior change.

2) Use RunConfig in one place in disco.py (or main entrypoint)
In disco.py (or whichever file currently wires these values together):

Import the new dataclass:

from pathlib import Path
from disco.config import RunConfig

Where it currently has something like:

output_dir = Path(args.output_dir)
device = args.device
seed = args.seed
run_diffusion(output_dir=output_dir, device=device, seed=seed, ...)

replace only this local wiring with:

cfg = RunConfig(
    output_dir=Path(args.output_dir),
    device=args.device,
    seed=args.seed,
)

run_diffusion(
    output_dir=cfg.output_dir,
    device=cfg.device,
    seed=cfg.seed,
    ...
)

This keeps the change very small and behavior-identical, while clearly moving the design toward a centralized settings object for this issue.

[vecnode]

Migrating runtime locals + run_args to RunConfig