v1.3.1

Changed

• Make torchfits.read fully discoverable with explicit keyword-only options
  and remove ineffective cache-policy objects, ignored tuning arguments, and
  private compatibility exports.
• Make every CLI command self-describing, add root --version, register the
  supported transforms explicitly, and reject inapplicable or conflicting
  options before touching output files.

Fixed

• Reject overflowing, truncated, or out-of-bounds binary, ASCII, and
  variable-length table geometry before allocation or I/O while preserving
  valid empty tables.
• Preserve every normalized option in multi-file reads and keep explicit mode,
  cache, scaling, dtype, row-window, and device behavior consistent across
  single-file and batch paths.
• Keep compression lossless by default, validate codec/dtype combinations,
  compare repeated header cards in order, complete multi-block remote headers,
  enforce probe deadlines, and prevent case or Unicode-equivalent output
  collisions.
• Build self-contained source archives and validate the exact wheel matrix,
  metadata, licenses, native linkage, macOS deployment floor, release-candidate
  provenance, archive safety, and artifact digests before publication.

Performance

• Avoid redundant floating-point NaN masks and unnecessary table-read dispatch
  work without benchmark-specific routing or sampling heuristics.

Documentation

• Synchronize the public API, CLI recipes, examples, supported Python/PyTorch
  lane, platform requirements, and guarded release procedure with the shipped
  implementation.

What's Changed

• Post-1.3.0 fixes: cache docs, keyword-only TableHDU, and an rc gate before PyPI by @sfabbro in #263
• release: torchfits 1.3.1rc1 by @sfabbro in #265
• release: prepare torchfits 1.3.1rc2 by @sfabbro in #266
• fix(release): accept compound manylinux tags by @sfabbro in #267
• release: torchfits 1.3.1 by @sfabbro in #268

Full Changelog: v1.3.0...v1.3.1