feat(json): emit Decimals as exact number literals (frontend-decimal epic, step A)

[robcohen]

Step A of the frontend-decimal epic — closes the wire half of C2 and fixes R8, and it landed far smaller than the original "regenerate 53 snapshots" estimate.

What

The JSON provider serialized every Decimal via float() (core/charts.py), so monetary values crossed the wire as lossy float64. Switch dumps to simplejson with use_decimal=True → Decimals emit as exact number literals (88.571428571428571428571428571 instead of 88.57142857142857).

• Exact wire — a decimal-aware API consumer recovers the exact ledger value; Python's lossy float() intermediate is gone.
• R8 fixed — custom/budget amount values are now exact, consistent with Balance/Price (which already stringified).
• Transparent to the frontend — exact literals are valid JSON numbers, so JSON.parse still reads them (float64 for display, same as today). No frontend change, no snapshot churn (the snapshot fixture re-parses to float before comparing).

Scope & verification

Diff is 3 files (charts.py, pyproject.toml, uv.lock) + the boundary test. Full py suite 100% coverage, 544 passed; JS 86 pass; just mypy clean.

test_number_precision_boundary rewritten: every value now crosses the wire exactly (the old high-precision xfail flips to passing), with one test documenting the residual — a plain JSON.parse/json.loads read still narrows to float64.

To weigh

• New core dependency: simplejson (+ types-simplejson stub). It's the clean way to emit exact Decimal literals (stdlib json can't — its C encoder ignores a float subclass's __repr__). loads stays stdlib json, so request parsing is unaffected.
• Step B (exact numbers displayed in the browser) is the remaining, larger part: a frontend decimal library on the read side. Lower value since the UI already rounds normal values for display; worth doing only if high-precision display matters.

🤖 Generated with Claude Code