json: pin the measure-then-write invariant with a test

[TurkeyMan]

Why

write_json() is designed to be called twice: once with a null buffer to measure, once to fill a buffer of exactly that size. The sync JSON encoder does this for every value it puts on the wire. If the two passes ever disagree, the caller emits whatever was already in the tail of its buffer — char.init, i.e. 0xFF — inside an otherwise well-formed frame, and the peer reports a JSON parse error with nothing to diagnose it from.

This came out of chasing intermittent malformed /sync frames from an ESP32-S3 board. The theory was a measure/write mismatch; the test disproves it for every shape that path encodes, which is worth keeping so the next person doesn't re-derive it.

What

Tests only — no production code changed, because none needed changing. The guards that came out of the same investigation are in open-watt/openwatt#700.

One unittest block, one check() helper, two sets of inputs:

• Quantity scalars — the unit spellings the field actually uses (W, kW, V, A, %, Wh, Hz, /s, /hr, …) crossed with 0, negatives, denormals, 1e100, -0.0, NaN and ±inf, plus an integer-valued quantity. Values the encoder refuses outright (negative measure) are skipped rather than asserted — refusing is legitimate, callers must handle it.
• Composites — nested maps, nested arrays, escapes and control characters, large and small numbers, empty object/array/string; each at indents 0/1/2/4, both dense and pretty.

Every case asserts the measured length is positive and in range, that the write pass returns exactly the measured length, and that no char.init byte survives anywhere in the written range.

Verification

make CONFIG=unittest — 74/74 modules pass.