v0.8.3

Correctness fixes for the OpenDSS writer and the BMOPF reader. No API, C ABI, or schema version changes: PIO_ABI_VERSION stays 4, PIO_DIST_ABI_VERSION stays 1, and .pio.json stays at schema 0.2.1. Two behaviors change in ways a consumer can observe, both below.

A center tapped service exports OpenDSS that solves to the right answer. Reported in eigenergy/PowerIO.jl#79 against three 19 kV SWER feeders. A dss node list is positional, the phase conductors first and the return last, and a center tapped service maps as [p1, n, p2]. The writer emitted one Load record over that map, so the engine discarded the conductor it could not address: on the reported network the load drew 0.652 kW of its stated 2.608 and the second leg floated to 1.35 pu, converged, with nothing in the warnings. Such a load now emits one single phase Load per leg. The split keys on the conductor count rather than on unequal per phase power, which a center tapped consumer does not have, and the return conductor is located from the bus grounding rather than assumed last. A terminal map longer than the record's conductor count now warns, the mirror of the short map warning that already existed.

Three more silent paths from the same report: a three winding star whose third arm is zero cannot be solved by OpenDSS, which collapses the secondary legs to about half voltage instead, so the writer substitutes the split from the OpenDSS center tap example and says so; the BMOPF and PMD readers warn when a document states no frequency and carries line susceptance, since the 60 Hz default costs a 50 Hz feeder about a fifth of its line charging; and a center_tap v_nom_to at about twice the secondary bus's own phase to neutral band warns, that being the full span rather than the per leg voltage the convention asks for.

A malformed BMOPF numeric field is refused rather than read as NaN. bmopf::read mapped every numeric field through as_f64().unwrap_or(f64::NAN), so a string, a null, an object, or an array holding one of those became NaN and the parse continued silently. Schema 0.1.0 spells no null anywhere and types the bounds and ratings as nonnegative_number, so every value that reached it was already invalid — and a NaN bound serializes into .pio.json as null, which the payload reader restores as ±Inf, an explicit "no limit" the source never stated. Each such field is now an Error finding carrying its JSON pointer, so powerio convert and powerio package exit nonzero on a document that used to parse quietly. The value itself still reads as NaN: telling absent from invalid per field is a real feature and it waits for the typed readers of #293.

Python. The MCP server is inside the ruff and mypy gates (#297); the exclusions #285 added are gone.

Repository. The normalize pass no longer clones the element fields it immediately overwrites, on a pass that runs before every solve.

What's Changed

• py: bring the MCP server under the ruff and mypy gates by @samtalki in #305
• core: stop cloning the fields normalize immediately overwrites by @samtalki in #306
• dist: split a center tapped load and repair a degenerate star by @samtalki in #303
• dist: refuse a malformed BMOPF numeric instead of reading NaN by @samtalki in #304
• Release 0.8.3 by @samtalki in #308

Full Changelog: v0.8.2...v0.8.3