v1.12.0

AnnData depth (CSC sparse, uns, labelled X), data-correctness fixes across
the CSV / display / numeric paths, security hardening, and broad RAM/throughput
work on the streaming and TUI paths.

Added

• AnnData uns decoding — a new key/value tab surfaces the unstructured
  annotations: scalars, strings and short arrays show their values, nested dicts
  flatten with dotted keys (pca.variance_ratio), and an encoded sub-object
  (dataframe / categorical) shows its encoding-type. Previously skipped.
• AnnData CSC sparse X preview — csc_matrix matrices now densify to a
  rows × columns preview just like CSR (the compressed axis is columns, not
  rows); they previously showed only a "not implemented" note.
• Labelled AnnData X preview — the sparse / dense X preview names its
  value columns by the var index (gene names) and prepends the obs index
  (cell barcodes) as a row-label column, instead of generic col0 / row numbers.

Fixed

• Leading-zero IDs in CSV/TSV (and spreadsheet / markdown / HTML tables) are
  preserved — a column of values like 007 is read as text instead of being
  inferred as the integer 7. Scientific notation stays numeric.
• Wide and combining characters align. display_width now measures terminal
  columns (CJK / fullwidth / emoji = 2, combining / zero-width = 0) instead of
  counting codepoints, so tables with such data no longer skew; truncation cuts
  on a column boundary.
• Hostile / malformed files are handled gracefully — AnnData sparse shape
  and DataFrame column lengths are validated against the real dataset extents
  (no out-of-bounds read).
• Date / timestamp / decimal columns are read as numeric values, so
  --describe min/max/mean and --heatmap work on them; --describe now
  summarises the whole table rather than the first chunk.
• Genomic / format correctness — Parquet region pruning indexes column
  statistics by Parquet leaf (not Arrow field); pileup honours reference skips
  (CIGAR N) like samtools; mid-stream htslib read errors in pileup / BCF
  region mode are surfaced instead of silently truncating; region mode reports
  exact post-filter row counts; CSV headers of nan / inf / hex are no longer
  mistaken for headerless numeric data.
• TUI / terminal robustness — the terminal is restored on
  SIGINT/SIGTERM/SIGHUP via an async-signal-safe handler; markdown rendering
  strips terminal control bytes; truncate never splits a multibyte codepoint;
  a missing flag argument reports the specific flag instead of "Unknown option".

Performance

• Bounded RAM on streaming sources — DelimitedSource and the other streaming
  readers evict old chunks behind a retention window instead of retaining every
  decoded batch; SQLite computes COUNT(*) lazily (not at open); generic HDF5
  1-D datasets and wide NPZ arrays preview a capped head; delimited preamble
  stripping is buffered. TUI redraws format each visible cell once and reuse the
  LRU chunk cache during sorted search.

Security

• SQLite identifier quoting escapes embedded double quotes, so a table named
  a"b no longer produces malformed / injectable SQL.
• NPZ allocation hint — the zip central-directory uncompressed_size (which
  is attacker-controllable) is no longer passed straight to reserve(), so a
  tiny crafted entry can't force a multi-GB allocation.

Build / CI

• KDE plugins stop including KDEInstallDirs, silencing the ECM directory
  warning in the GUI build.