write_bam_record was the single biggest hot path after the v4.4.3 variant
work — 28 M Seq.__getitem__ calls per 185 k-read run, plus a per-base list
comp for the quality string and 9 small struct.pack calls per record.
Replace with bulk numpy / single struct.pack operations:

- Sequence encoding: replace the per-byte Python loop
    for i in range(encoded_len):
        encoded_seq.extend(pack('<B',
            (SEQ_PACKED[alt_sequence[2*i]] << 4) +
            SEQ_PACKED[alt_sequence[2*i+1]]))
  with a vectorized lookup through a 256-entry table:
    codes = SEQ_PACKED_LOOKUP[np.frombuffer(seq_str.encode('ascii'), ...)]
    encoded_seq = ((codes[0::2] << 4) | codes[1::2]).tobytes()
  Eliminates the 28 M Biopython Seq.__getitem__ calls that dominated cumtime.

- Quality encoding: replace
    encoded_qual = "".join([chr(n) for n in read.quality_array]).encode(...)
  with one `np.asarray(quality_array, dtype=np.uint8).tobytes()` call.

- CIGAR ops: pack all in a single `struct.pack(f'<{n}I', ...)` instead of
  one pack per op + bytearray.extend. Also drop the regex CIGAR re-parse in
  favour of a single linear scan over the CIGAR string (no regex import
  needed; `re` removed).

- BAM record header: one `struct.pack('<iiiIIiiii', ...)` for the nine
  fixed-size fields instead of nine separate pack calls glued with `+`.

Effect (clean machine, ecoli 10x 4-thread untimed):
  SE wall: 1:35.3 (v4.4.3) -> 1:07.1  (1.42x faster, -30%)
  PE wall: 1:34.5 (v4.4.3) -> 1:05.9  (1.43x faster, -30%)
  SE CPU%: 348 -> 364   PE CPU%: 359 -> 378   (workers near-saturated)

Single-thread cProfile of write_bam_record itself:
  self time:   12.0 s -> 3.2 s   (3.75x)
  cumtime:     58.1 s -> 8.1 s   (7.2x)

Combined with the v4.4.4 variant-generation fixes (committed separately) and
the v4.4.3 architectural work, NEAT is now 13.3-13.4x faster than v4.4.2 on
the ecoli benchmark, with strictly coordinate-sorted output and bounded
per-worker memory.

All 599 tests pass. Bumps version to 4.4.4 for the release.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Records the three hot-path fixes in this release with before/after numbers:
  - ContigVariants overlap-check O(1) via position-bucket dicts
  - map_local_trinuc_bias vectorization (16x)
  - write_bam_record vectorization (7.2x cumtime)

Total release effect on ecoli 10x 4-thread:
  SE 1:35 -> 1:07 (1.42x)
  PE 1:34 -> 1:06 (1.43x)
13.3-13.4x faster than v4.4.2 cumulative.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Low-risk audit items applied across the README:

- Prerequisites: poetry pinned to >= 2.0 (was 1.3.*, no longer accurate against
  pyproject.toml's poetry-core>=2.0 requirement). Added version constraints
  for numpy (>=2.2) and pysam (>=0.23) to match pyproject.
- Removed the OS-support contradiction. The Prerequisites section already
  notes "NEAT assumes a Linux environment ... install WSL on other OSes"; the
  installation section now agrees ("Linux tested; macOS may work; Windows via
  WSL") instead of contradicting it with the old "supports MacOS, Windows,
  and Linux" claim.
- Usage table: corrected `mode`/`size` to the actual config keys
  `parallel_mode`/`parallel_block_size`. Updated descriptions for the new
  defaults (`parallel_mode: size` for threads>1, forced to `contig` for
  threads==1; `parallel_block_size: 0` = auto-tune).
- Estimated-runtimes banner: updated to point at v4.4.4 numbers (~13× faster
  than v4.4.2 on ecoli) and to reference both v4.4.3 and v4.4.4 ChangeLog
  entries.
- Estimated-runtimes config snippets: same `mode`/`size` → `parallel_mode`/
  `parallel_block_size` correction.
- "Insert specific variants" example: dropped the `-v`/`-M` CLI flag prose
  (those flags were removed when NEAT switched to config-file-only).
- Three remaining Examples sections (Insert variants, Single-end, Large
  single-end): matched the fence formatting pattern (separate yml + bash
  blocks, "Contents of neat_config.yml:" label) that was already applied to
  the first two examples.
- Long-read caveat: reworded "not yet implemented in NEAT 4.0" to "planned
  for a separate future release; not yet fully validated in NEAT 4.4".
- Removed the `neat vcf_compare` section entirely. The subcommand does not
  exist (confirmed via `neat --help`) and the section was documenting an
  unimplemented interface.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…om threads

`parallel_mode` was a YAML option with two values, `'size'` and `'contig'`,
which never actually let the user change behavior:

  - When `threads == 1`, Options force-set it to `'contig'` regardless of
    user input.
  - When `threads > 1`, `'size'` was the default; `'contig'` would have
    yielded one worker per reference contig, but `chunk_record` in
    split_inputs already produces a single chunk for contigs ≤ chunk size,
    so `'size'` is a strict behavioral superset.

Removing it cleans up dead config and eliminates a footgun (a user setting
`parallel_mode: contig` with `threads: 8` thinking they'd get one-worker-
per-contig would in fact just disable sub-contig splitting on large
contigs and silently reduce parallelism).

Changes:

- options.py: drop `parallel_mode` from the Options __init__ signature,
  schema, and stored state. Replace the post-parse log block with a
  threads-keyed one. Add a `DEPRECATED_KEYS` table and a deprecation
  warning in `read_yaml` so old configs with `parallel_mode: ...` keep
  parsing (the key is silently ignored after a one-line warning).
- split_inputs.py: branch on `options.threads == 1` instead of
  `options.parallel_mode == "contig"` for one-chunk-per-contig output.
- runner.py: simplify the auto-tune gate; drop the
  `options.parallel_mode == "size"` predicate (always true when
  `threads > 1`).
- config_template/template_neat_config.yml,
  config_template/simple_template.yml: drop the `parallel_mode: .` line
  and tighten the comment for `parallel_block_size`.
- README.md: drop the `parallel_mode` table row; clean up references in
  the Estimated runtimes section and HPC examples; tighten the
  parallel_block_size description.
- tests/test_models/test_split_inputs.py: the `mode=` parameter in the
  test helper now maps to `opts.threads` (1 for contig-style, 4 for
  size-style) — callsites unchanged.
- tests/test_read_simulator/test_options.py:
  - drop `parallel_mode` lines from the embedded YAML samples
  - drop the `assert opts.parallel_mode == "contig"` default-values
    assertion
  - swap the `parallel_mode` example in the choice-validator tests for a
    synthetic key (`_test_choice_key`) since no production option uses
    the `choice` validator anymore
  - replace `test_log_configuration_threads_one_forces_contig` (which
    was checking a force-override that no longer exists) with
    `test_read_yaml_deprecated_parallel_mode_warns` (verifies the
    deprecation warning + silent-ignore path works end-to-end via
    `Options.from_cli`)

End-to-end check: an old YAML with `parallel_mode: size` still runs to
completion and emits a single warning line:
  WARNING: `parallel_mode` is deprecated and ignored; splitting strategy
  is now derived from `threads`.

All 599 tests pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The removal is a YAML-schema cleanup; existing configs keep working via a
deprecation warning path. Behavior is unchanged — parallel_mode never
actually let users change anything under the old logic.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Per-chunk read counts in cover_dataset were scaled by chunk_mean/max_weight
when a non-uniform gc_model was configured, making `coverage` the peak
coverage at the GC bias maximum rather than the documented genome-wide
average. The fix precomputes the genome-wide mean weight once at the
runner level and divides per-chunk reads by that, so the chunk-scaling
ratio averages to 1.0 and total reads match coverage * genome_length / read_len.

- New compute_genome_wide_gc_mean_weight helper (vectorized, reuses the
  same prefix-sum windowing as cover_dataset; uniform models short-circuit).
- Runner stores the result on options.gc_global_mean_weight so workers
  don't redundantly scan the reference.
- cover_dataset falls back to chunk-local mean (no rescaling) when the
  field is unset, preserving single-chunk unit-test behavior.

Drive-by: drop dead MarkovQualityModel stub and stale TODO in
test_contig_variants.py (the remove_variant fix landed separately).

ChangeLog rewritten as a single v4.5.0 entry covering the perf work
already shipped in v4.4.4, the parallel_mode removal, and this fix.
All 603 tests pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Five comments were either factually wrong, idle speculation, or dead
work-in-progress with no path back into the code:

- model_sequencing_error/utils.py: drop "TODO implement plotting" — plt
  is in fact used heavily later in the file. Remove the commented-out
  shape_curves.append() + the 8-line shape_curves DataFrame scaffolding;
  both referenced pandas/seaborn that aren't imported and have sat dead
  for years.
- model_fragment_lengths/utils.py: drop speculative docstring TODO about
  filter_lengths "maybe" being useful elsewhere.
- common/ploid_functions.py: drop "TODO may need to improve the modeling
  for polyploid, maybe" — pure musing, not actionable as written.

No behavior change; 603 tests still pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Both lists were built up in gen_mut_model/runner.py but never read or
returned anywhere in the codebase — scaffolding for two features that
were never wired into the saved mutation model:

- common_variants: 95th-percentile-frequency variants from the input VCF.
- high_mut_regions: variant-density-clustered hotspots, top 3% by density,
  with overlapping regions merged.

The design intent (pipe both into the saved mutation model so the
simulator can bias generation accordingly) is now captured in #295. If
the work is revived, the cluster_list and find_caf helpers are still
available in gen_mut_model/utils.py.

Collateral cleanup: drop the now-unused vcf_common dict and the
cluster_list / find_caf imports from runner.py.

No behavior change in the produced model; 603 tests still pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The release that adds the GC bias coverage fix, the follow-up perf pass,
and the parallel_mode removal will be tagged 4.4.4 instead of 4.5.0.

The GC bias fix does change output bytes when a non-uniform gc_model is
configured, but the user-facing config has always documented `coverage`
as the genome-wide average — the prior behavior was the bug. Shipping
as a patch release is consistent with framing this as a fix rather than
a deliberate behavior change.

Updates:
- pyproject.toml: 4.5.0 -> 4.4.4
- ChangeLog.md: rename v4.5.0 entry -> v4.4.4; benchmark column relabeled;
  drop the "reason for the minor version bump" sentence and replace it
  with concrete coverage-shortfall numbers (~70 % delivered pre-fix).
- README v4.4.4 references are already correct; no edit needed.

603 tests still pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>