Vendor codec 2.65.0 (from vaptvupt-codec tag v2.65.0), re-applying the two
in-tree audit patches on top: the ANS decode safe-zone 2*SAFEZONE_MAX_RUN
reserve (heap-overflow guard, not yet upstream) and the AVX2 offset-read
bound in vv_decoder.c.
Two settings in the integration layer were leaving most of the codec's ratio
on the table:
* vaptvupt_api.c forced opts.format_v2=1 for balanced+extreme. Since codec
v2.61.0 that routes text through the binary/greedy path and HALVES the
extreme-mode text ratio (codec-level 7.6x -> 3.7x). The codec auto-enables
format_v2 for binary-detected input on its own, so stop forcing it: text
keeps the optimal parser, binary still gets v2.
* auto_block_size() capped the extreme block at 512 KiB. The block IS the
codec's LZ window, so the 'large-window extreme' parser could never match
past 512 KiB. Scale block size with level (128 KiB fast -> 8 MiB extreme).
Because block size also sets --dedup granularity (a large block rarely
finds a byte-exact duplicate), --dedup now overrides to a small 256 KiB
block so block-level dedup still works.
Measured, level 9 extreme: text 3.77x->5.98x (+58%), logs 7.21x->9.07x (+26%),
json 8.25x->9.38x, source 4.93x->5.63x. Wire format unchanged (v1.6); 5.0.0
and 5.1.0 archives interoperate both directions (verified, all modes). Bump
ZUPT_CODEC_RELEASE to 2.65.0 and ZUPT_VERSION_STRING to 5.1.0.