v0.8.1

Text writer hardening, JSON reader fidelity, a document version report
over the C ABI, and a release gate that tests the Julia binding before any
binary ships. No breaking changes over v0.8.0.

Security hardening. The psse, pslf, powerworld, and OpenDSS writers replace a
line terminator inside any quoted or free-text field. A name that held a
\n or \r ended the record and made the rest of the text parse as new
records, so a crafted case name, DC line label, circuit id, or bus name
could forge whole records in the written file. Five such paths are now
closed, each with a test that reads the written text back and counts the
records. Upgrade if you write text formats from names you do not control.

• The two C JSON report entry points, pio_last_error_json and
  pio_schema_versions_json, now run inside the panic guard the other
  entry points use. A panic in either one crossed the C boundary, which is
  undefined behavior.
• build_ybus refuses a case whose base MVA is zero or nonfinite. Such a
  case wrote a matrix of NaN and the CLI exited 0.
• The auto sensitivity solver reads the real matrix shape (branches by
  buses) instead of the reduced dimension alone, and holds the dense path
  to a 2 GiB memory budget. The dense threshold moves from 512 to 8192,
  because a dense solve is faster than conjugate gradient well past 512
  buses. Set SensitivityOptions::auto_dense_threshold to keep the old
  value. PTDF and LODF results are unchanged: both paths were verified
  bit-identical on case30.
• The psse field splitter, the powerworld auxiliary token splitter, and
  the pandapower split-frame reader reuse their buffers and move rows
  instead of copying them. Reader output is unchanged.
• New pio_schema_versions_json C entry point reports the schema version
  of every document format the library speaks (.pio.json, Arrow, the
  distribution capability document, and the BMOPF vintage). A key is
  null when the owning feature is not compiled in. PIO_ABI_VERSION
  does not cover document formats, so a binding that mirrors one of these
  versions can now read it from the library and refuse a mismatch at load
  or pin time instead of finding it downstream (#270, query half).
• release-binaries.yml gates every tag on PowerIO.jl: the workflow
  builds the tag, runs the binding's suite against it, and produces no
  tarballs and no draft release on failure. A planned binding break
  merges the paired PowerIO.jl change first, then re-runs the workflow.
• The retired schema documents under docs/schema/ are pinned by a test.
  A .pio.json written before v0.8.0 declares those URLs, so they stay
  published even though the reader no longer accepts that lineage.
• The egret and pandapower readers keep unrecognized element fields as
  extras instead of dropping them silently, matching the PowerModels
  reader. A powerio-written file still reads back extras-free (#263).
• The PowerModels reader reports the taps it discards: a branch with an
  off-nominal tap but no transformer: true flag reads as a line, and
  one aggregated warning now names the branches and the total. The
  inference rule itself is unchanged.
• .pio.json validation diagnoses duplicate payload uids directly
  (VALIDATE.BALANCED.PAYLOAD_IDENTITY, a new always-present
  balanced.payload_identity pass) instead of leaving the ambiguity to
  surface later as a failed operating-point reference.
• The text-only C conversion entry points warn when a distribution
  writer produced a companion file they cannot return, naming the file:
  an OpenDSS deck referencing a Buscoords CSV the caller never received
  no longer fails to compile with nothing to explain why.
• summary, package, convert without --from, and geo extract|apply read and classify a .json once instead of twice. The
  distribution reader's own document rule still applies on that path, so
  a JSON that is not a distribution case is refused as before.
• Python Network.ptdf() / lodf() route through the auto sensitivity
  solver (dense below the reduced-dimension threshold, iterative
  conjugate gradient above it), matching the CLI sensitivities
  command. Both take solver="auto"|"dense"|"iterative". On a case above
  the threshold results move from exact-dense to iterative-CG at a 1e-10
  relative residual (#273).

What's Changed

• Report schema versions over the C ABI and gate releases on PowerIO.jl by @samtalki in #280
• docs: make AGENTS.md and languages.md match the tree by @samtalki in #282
• py: route ptdf/lodf through the auto sensitivity solver by @samtalki in #281
• format: capture unknown JSON fields; warn on discarded taps; single read by @samtalki in #287
• Close the record-injection seams; fix the auto solver policy and reader allocation churn by @samtalki in #295
• Release 0.8.1 by @samtalki in #290

Full Changelog: v0.8.0...v0.8.1