v0.5.2

Added

• fqxv.estimate() accepts paired-end (and other grouped) inputs. Pass a list
  or tuple of sources — paired mates, 10x R1/R2/I1/I2, sharded lanes — and each is
  sampled with a split of the sample_reads cap, with the per-stream sizes summed
  into one aggregate Estimate, matching fqxv compress … --estimate byte for
  byte. Single-source behavior is unchanged. Estimate also gains a platform
  field ("illumina", "nanopore", "pacbio", "mgi", "unknown").
• Cross-platform guard on grouped estimates. A group whose inputs resolve to
  two different known platforms is now a hard error in both the Python API and
  the CLI's --estimate, instead of silently summing two differently-calibrated
  numbers. An Unknown input (SRA-renamed, no content signal) stays compatible
  with anything, so real SRA data never trips a false positive.

Changed

• New fqxv-align crate. The two edit-distance alignment implementations —
  banded Needleman-Wunsch (AVX2 + scalar) and the wavefront aligner — moved out of
  fqxv-lroverlap into a zero-dependency leaf crate, so an external consumer can
  reach them without pulling in the whole long-read codec cone. The move is a pure
  rename: fqxv-lroverlap re-exports all seven public names at their existing
  paths, and its public surface is unchanged. fqxv-lroverlap is now
  #![forbid(unsafe_code)], since the AVX2 backend that motivated the weaker
  deny has left the crate.
• Dependency bumps: noodles-bgzf 0.48 → 0.49, clap 4.6.2 → 4.6.3,
  xxhash-rust 0.8.17 → 0.8.18, serde_json 1.0.150 → 1.0.151.