fix: merit NaN guard + hygiene batch

[corriander]

Summary

A "fixes & hygiene" batch, split into six logical commits.

Task 1 — merit NaN guard

CODA.satisfaction divided total merit by total correlation per requirement. A requirement whose matrix row is entirely CODANull has zero total correlation, so the division produced a silent nan that poisoned .merit. Now such requirements are detected and a clear ValueError is raised naming each offender (e.g. Requirements have no relationships defined: 'Stiffness', 'Weight'), so the scorer fails loud instead of silently. The FIXME is gone. The coda CLI command catches this and reports it via the rich console with a non-zero exit code. Tests added for single/multiple unlinked requirements; the fully-linked case-study values are unchanged.

Task 2 — hygiene

• docs: from_google_sheets → from_google_sheet (real method) in the gsheets guide; phantom BinWMGSheet → real GSheetBinWM in the google sheets API doc.
• build: dependency typer[all] → typer (same version constraint) — the all extra no longer exists upstream and uv sync warned. uv.lock was deliberately regenerated with one explicit uv lock; the diff is minimal (just the extra removed from vdd metadata, no dependency churn).
• refactor(common): removed the vestigial py2/3 ABC = abc.ABCMeta(...) shim (vdd/common/abstract.py) in favour of stdlib abc.ABC; project requires >=3.11. Usages in vdd/common/__init__.py and vdd/common/io.py updated.
• fix(coda): compare: CODA.compare(other) returned an elementwise ndarray, not a bool. It now returns True iff shapes match and all corresponding relationships compare equal, reusing CODARelationship.__eq__ (guarded by a shape check first). Docstring and tests added.
• fix(coda): relationship equality (review follow-up): CODARelationship.__eq__ compared only (correlation, target), so CODAMaximise(0.9, 5) == CODAMinimise(0.9, 5) was True and a zero-correlation relationship compared equal to CODANull — meaning compare could report two models equivalent with a relationship's direction flipped. Equality now requires the concrete types to match, and returns NotImplemented for non-relationship operands instead of raising AttributeError. CODAOptimise's tolerance extension is preserved on top.

Testing

uv run --frozen pytest — 165 passed (154 baseline + 11 new tests). CLI happy and error paths verified end-to-end.

Judgement calls

• compare semantics: compare relies on CODARelationship.__eq__ rather than re-deriving type tuples; per review, the equality contract itself was fixed to be type-aware (see above) so the reuse is now sound. A shape guard avoids ndarray broadcasting surprises.
• CLI error handling: cheap to add, so the coda command now wraps merit/satisfaction access and surfaces the ValueError as a red console message + typer.Exit(code=1) rather than a traceback.

🤖 Generated with Claude Code