feat: native JSON model i/o and BinWM→CODA glue

[corriander]

Summary

Adds stdlib-only (json) native serialisation for CODA models and Binary Weighting Matrices, and glue to seed a CODA model directly from a BinWM. No new dependencies.

CODA JSON

Flat, human-readable, git-diff-friendly schema:

{
  "requirements": [{"name": "...", "weight": 0.2, "normalise": true}],
  "characteristics": [{"name": "...", "limits": [lo, hi], "value": 24}],
  "relationships": [
    {"requirement": "...", "characteristic": "...",
     "type": "min|max|opt", "correlation": 0.9,
     "target": 0.5, "tolerance": null}
  ]
}

• CODA.to_dict() / from_dict() — dict form.
• CODA.to_json(path=None) — returns a string when path is omitted, else writes the file.
• CODA.read_json(path) — classmethod mirroring read_excel.
• Only non-null relationships are emitted. limits/value may be None; tolerance is None for non-optimising relationships.
• Requirements store the raw weight (base_weight for self-normalising requirements, otherwise the pre-normalised weight) plus a normalise flag, so from_dict replays normalisation on the target instance via the public add_* API — no element objects are copied across instances (their weights are context-dependent).

BinWM JSON

BinWM.to_dict() / from_dict() / to_json() / read_json() using the exact shape of the existing test fixtures:

{"requirements": ["..."], "binary_matrix": [[0, 0, 1], ...]}

The bundled case__*.json fixtures are therefore valid interchange files (a test loads one directly through read_json).

BinWM → CODA glue

CODA.add_requirements_from(binwm) adds one requirement per BinWM requirement, using its score as the weight. BinWM.score sums to unity by construction, so scores are added with normalise=True (self-normalising requirements) — this keeps floating-point drift around unity from tripping the combined-weight guard and lets further requirements be mixed in afterwards with automatic re-normalisation.

Tests

14 new tests (168 total, was 154). CODA round-trips through dict, JSON string and file for the bicycle-wheel case study (merit/satisfaction/correlation/weights and all names/limits/values/relationship params asserted), plus optimise-tolerance + unset-value handling and pre-normalised weights. BinWM round-trips including a fixture loaded directly. README updated with a JSON usage example.

🤖 Generated with Claude Code