[matrixarkai]

Every row in the allocation report stores its value raw. The fixtures fill
with incompressible bytes on purpose — a compressible value shrinks to a
forty-four byte block and stops exercising the tier, a trap that has cost this
work two measurements already.

The consequence is that the compressor has been running on every one of those
writes and having its output discarded, and the decompressor has never run at
all.

a 1 KiB value	allocations	bytes
get, promoting a compressed block	14.00	153,389.0
get, promoting a raw block	12.00	24,310.0
put, compressing a block (pure-Rust store)	25.86	49,498.2
put, compressing a block (the other backend)	10.83	48,342.4

A read of a compressed block allocates 153,389 bytes to return 1,024 —
six times the same read when the block is stored raw. The figure is identical
on both feature arms to the byte, so none of it is the store: it is
zstd::stream::decode_all building a decompression context and its window on
every call.

The write is the same thing in the other direction. A write that keeps its
compressed form pays 49,498 bytes; a write whose value does not compress
pays 58,224 for a payload of the same size, because the compressor runs there
too and the result is thrown away. The difference between those two rows is
not the compressor — it is the block that gets stored.

Both rows assert the regime they are named for

The read checks that nine in ten of its reads reached a lower tier, and that
nine in ten of those were served from a block with the compressed codec byte
set. If the fixture's value ever stopped compressing, the row fails rather
than quietly becoming a second copy of the raw one.

The bounds are where the codec is today, not an endorsement of it. A reused
context would move both figures, and these come down with it.

Gates on both feature arms: cargo fmt --check, clippy --all-targets -D warnings, both suites, the vocabulary check, and both reports with
--require-passed.

🤖 Generated with Claude Code