v0.9.0

Install: pip install oxihipo==0.9.0

Added

• Compression is a (codec, layout) pair, not a flat list of six named
  combinations: Codec::{None,Lz4,Lz4Hc,Gzip,Zstd} x
  Layout::{PerChunk,PerBank,PerColumn}. All 15 pairs have a wire tag and
  round-trip. The six historical names survive as associated constants, so
  Compression::Lz4PerColumn is still valid source at all 229 call sites and
  still means exactly what it meant (Lz4Hc x PerColumn).

• Zstandard, levels 1-6 via Compression::with_zstd_level. The level is a
  writer-side knob and never reaches the wire — one tag decodes them all,
  unlike LZ4/LZ4-HC which burn two. On a 248 MB real CLAS12 file,
  Zstd x PerColumn is 2.22x smaller and scans in 20.9 ms, against
  2.03x/28.2 ms for Lz4Hc x PerColumn and 2.32x/21.9 ms for
  Gzip x PerColumn — but writes in 0.69 s where gzip takes 2.52 s and LZ4-HC
  7.66 s.

  Tags 4 and 5, left poisoned when Lz4Chunked and Lz4ByBank v1 were
  removed in 0.x, are reused for Zstd. That is safe specifically because a
  zstd frame begins with the magic 0xFD2FB528: a stale file carrying one of
  those tags fails the frame check rather than decoding as something
  plausible. Tag 15 is the only one left unassigned.

  Only the six pairs that predate the matrix are readable by hipo-cpp and
  hipo-java; the other nine are oxihipo extensions those readers reject as
  an unknown tag. The split-record directory stays LZ4 for every layout so
  tags 6 and 7 remain byte-compatible.

• Python: compression= takes the pair too — "<codec>+<layout>", e.g.
  "zstd+percolumn" or "zstd6+perbank". A bare codec means perchunk, and
  the six older names still work and still mean the same thing
  ("lz4percolumn" is lz4hc+percolumn, not lz4+percolumn — the split
  codecs were always high-compression). An unknown name lists what is valid;
  a zstd level outside 1-6 is an error rather than a silent clamp.

• Compression prints as lz4hc+perbank / zstd3+percolumn — the same
  grammar the Python binding accepts, so format!("{c:?}") round-trips back
  through it. The derived Debug printed a struct dump containing {, } and
  :, which is noise in a log and illegal in a Windows filename.

Fixed

• read_columns handed back buffers carrying their growth slack. Assembly
  grew each column with extend_from_slice per record, so the result kept the
  last doubling's headroom — measured at 1.664x the payload on a real
  CLAS12 DST, retained for as long as the caller held it. That is the lifetime
  of the NumPy array for the Python binding, where into_pyarray moves the
  Vec across with the slack included. Now 1.000x.

  Every chunk is already in hand when the buffers are assembled, so the final
  length is known before any appending and the buffers are sized exactly up
  front. That is faster than the naive fix as well as smaller: a first version
  grew and then shrink_to_fit, which left best-of-15 unchanged but median ~5%
  worse on a 9-column read (the end-of-assembly realloc). Sizing up front
  removes both the doubling and the realloc — measured against the pre-session
  baseline, best-of 1.283 -> 1.229 s and median 1.606 -> 1.551 s.