v0.6.4

Fixes reading the ##ANCESTRY header FLARE actually writes. FLARE names
its ancestries on one line, keyed by name:

##ANCESTRY=<EAS=0,NAT=1,SAS=2,EUR=3,AFR=4>

v0.6.1 through v0.6.3 only understood a different spelling -- one
##ANCESTRY=<ID=0,Name=AFR> line per ancestry -- and refused anything else
with FLARE ##ANCESTRY header line lacks ID or Name. Since v0.6.2 also removed
--n-ancestries, there was no way to convert a real FLARE VCF on those
versions at all.

Both spellings are read now. If you hit that error, this release is the fix;
nothing about a prefix built by an earlier version changes, since the ancestry
codes were always taken from AN1/AN2 and were never renumbered.

---

Linux x86_64 static binaries: felixla and vcf_tbi_chunks.
Verify with the .sha256 file beside each one.

v0.6.3

Fixes an extract that does not finish on a real prefix. v0.6.2 dropped the
ancestry blocks no retained variant falls in, and copied the ones it kept by
seeking to each. Those handles carry a 16 MB buffer, so every seek spent a
16 MB refill to deliver one 200 KB block -- invisible on a small fixture, fatal
at scale.

Measured on chr22 of a 400,000-sample cohort, subsetting a 72 GB ancestry
payload to HapMap3's 40,846 sites and keeping 33,224 of 360,920 blocks:

v0.6.2   killed after 3 h, having written 5 GB of the 6.6 GB output
v0.6.3   about 21 min, which is what the same subset cost in v0.6.1

v0.6.3 walks the payload forward instead, reading through small gaps and
seeking only over large ones. Nothing about the output changes: the same
blocks are kept and the files are identical to what v0.6.2 would eventually
have produced.

If an extract under v0.6.2 appeared to hang, this is why. Nothing else in
v0.6.2 is affected -- the packing paths, the exports and the filters are
untouched.

---

Linux x86_64 static binaries: felixla and vcf_tbi_chunks.
Verify with the .sha256 file beside each one.

v0.6.2

Filtering variants now filters the ancestry blocks with them. A block that
no retained variant falls in is dropped, where v0.6.1 kept every block whatever
was filtered out of it. A prefix rebuilt from the same --extract or
--exclude therefore has fewer ancestry blocks, and --export-lai and
--export-local-admixture have correspondingly fewer rows.

This changes per-sample admixture proportions, because they are weighted by
the span the blocks cover and that span is now smaller. The difference is not
cosmetic. Take a sample that is EUR at both ends of a chromosome and AFR
across the middle, and extract only variants from the two ends:

v0.6.1   s1   AFR 0.600000   EUR 0.400000
v0.6.2   s1   AFR 0.000000   EUR 1.000000

v0.6.1 averaged in three blocks whose variants had all been filtered out.
Proportions computed from an extracted prefix under v0.6.1 should be
recomputed.

--export-global-admixture also warns now when the blocks leave a hole inside
a contig, which is what an extract leaves behind.

Extracting from a prefix with many ancestry blocks is much faster: the
per-variant block lookup was a linear scan over every block and is now a binary
search. A 20,000-variant, 15,000-block extract went from 0.82s to 0.14s.

---

Linux x86_64 static binaries: felixla and vcf_tbi_chunks.
Verify with the .sha256 file beside each one.

v0.6.1

FELIXla v0.6.1 Linux x86_64 static binaries: felixla and vcf_tbi_chunks.

v0.6.0

FELIXla v0.6.0 Linux x86_64 static binaries: felixla and vcf_tbi_chunks.

v0.5.0

FELIXla v0.5.0 Linux x86_64 static binaries: felixla and vcf_tbi_chunks.

v0.4.1

FELIXla v0.4.1 Linux x86_64 static binaries: felixla and vcf_tbi_chunks.

v0.4.0

FELIXla v0.4.0 Linux x86_64 static binaries: felixla and vcf_tbi_chunks.

v0.3.1

FELIXla v0.3.1 Linux x86_64 static binaries: felixla and vcf_tbi_chunks.

v0.3.0

FELIXla v0.3.0 Linux x86_64 static binaries: felixla and vcf_tbi_chunks.