Two accuracy defects in reader-facing benchmark output.
The persist charts' caption still said "Save: full index → .tv file,
calibration frozen". #474 removed automatic TQ+ calibration, so a default
index performs none during `save()` and there is nothing to freeze. The
claim shipped in four figures — `docs/{arm,x86}_persist_{st,mt}.svg` —
which are regenerated here. The insert captions lost the same claim
earlier; this was the last one.
`write_compression_chart` hardcoded its three rows while every other
chart in the file reads `results/`. The numbers happened to agree with
`compression.json`, but only by hand: a rerun that moved the JSON would
leave the shipped figure silently stale, which is the defect rather than
the current values. It now reads the JSON, and asserts the 2-bit and
4-bit rows of each dataset agree on `fp32_mb` — they describe the same
corpus, so a disagreement means one is stale.
Verified both directions: with the current data `compression.svg`
regenerates byte-identical, and perturbing a value in the JSON moves the
chart, so the coupling is real rather than decorative.
Closes#492
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Ryan Codrai
44ae40b93e