A second adversarial round (with empirical reproduction against the
vendored image 0.25) found the budget contract still had holes:
- The probe billed 4 bytes/px, but 16-bit formats allocate their
native layout during decode (8/px for RGBA16 PNG; TIFF stacks ~3x
codec scratch on top). In-budget 16-bit images were falsely refused
with self-contradictory notes, and the decode transient could exceed
the budget the tiers claimed to enforce. The probe now reads the
color type alongside the dimensions and bills the real peak (native
layout + RGBA where they coexist, plus the TIFF codec factor), so
refusals are honest and per-platform-correct, and the TIFF allowance
follows the same accounting.
- GIF: the probe reads the LOGICAL SCREEN, but a crafted frame may
declare itself arbitrarily larger (empirically: a 10x10 probed
screen with an 8192x8192 frame — 320 MB and 29 s from a
probe-claimed-400-byte image). The GIF frame buffer is now capped
to the probed size plus slack; frame bombs refuse like any other
oversize.
- Remote bodies on web streamed incrementally with the 64 MB cap
enforced DURING the read (array_buffer() materialized a hostile
chunked reply in full before any check could fire), aborting and
cancelling the reader the moment the cap is crossed.
- Degenerate raster entities (zero/non-finite size or vectors) no
longer produce NaN placement math that aborted whole PDF exports —
they resolve to the broken-reference placeholder.
- Same-name resolution now has a deterministic winner (newest
registration, via insertion order) instead of HashMap iteration
order; URL query strings no longer break by-name matching; the
image pick matches its refusal note by the quoted file name
(substring collisions like a.tif/2011.tif attributed errors to the
wrong file); a dropped remote outcome at the queue cap triggers one
unconditional catch-up re-resolution instead of waiting for another
event; the unknown-size refusal says so instead of printing 0x0;
and the build script fails cleanly when there are no fonts to
verify.
Review of the image wave (with empirical reproduction against the
vendored image 0.25) found the budget contract did not hold across
formats: the decoder's Limits reserve checks NATIVE bytes-per-pixel,
so a grayscale/RGB JPEG/PNG/BMP over the RGBA budget decoded anyway
and into_rgba8() then ballooned to up to 4x it — the tab death the
budgets exist to prevent. The probed dimensions already prove an
OverBudget image cannot fit, so it is now refused without decoding,
removing all dependence on per-decoder limit fidelity.
Conversely, image's TIFF codec stacks ~3x scratch buffers on top of
the reserve, so an in-budget TIFF (the flagship GIS scenario) was
refused with a self-contradictory note ('400 MB, over the 512 MB
limit'); the decoder allowance for TIFF is now 3x while tier math
stays on the real budget.
Also: the refusal message saturates instead of overflowing on crafted
headers, the image pick only matches the refusal note naming ITS file
(peek, not drain — other references' notes survive for their own
drains), and the Georgian Supplement range ends at U+2D2F (U+2D30.. is
Tifinagh, which ships no font).
Build tooling: the font files missing from dist was a trunk
finalize_dist failure (Windows file-lock during the dist wipe; staging
had them, the exit code was missed) — not a copy-dir skip; it did not
reproduce in two clean cycles. scripts/build-web.sh is now the
canonical wrapper (MSYS_NO_PATHCONV, trunk, font copy + byte
verification that hard-fails), wired into README, Trunk.toml, and the
Pages workflow.
Afonsop