fix(pdf): inflater fed in 16 KiB slices (browser memory bound), shared 16 MiB candidate budget, monotonic bounded endstream search, JSON object pre-check

This commit is contained in:
bchanot committed 2026-10-11 17:08:56 +02:00
1 parent 6adc38b027
commit 3a6e92625a
6 files changed
+310 -32

No files matched your search

+7 -4
View File
@@ -143,10 +143,13 @@ never lost because of its wrapper:
with its direct `/Length` or, failing that, the distance to `endstream`;
a `/Filespec` dictionary in clear whose `/EF /F` points at the stream
gives it a name. Candidates named `osteo-report*` come first, then file
order; Flate data is inflated by the platform inflater under a 16 MiB cap
(cancelled beyond it), and the first candidate whose text is a JSON
object with a numeric `envelopeVersion` or `version` wins. Object
streams, cross-reference streams and page contents are never inflated.
order; Flate data is fed to the platform inflater in 16 KiB slices, so
memory stays bounded in browsers too, under a 16 MiB cap (cancelled
beyond it); at most 16 MiB of candidate content is examined in all, a
candidate not opening on `{` is skipped unparsed, and the first one whose
text is a JSON object with a numeric `envelopeVersion` or `version`
wins. Object streams, cross-reference streams and page contents are never
inflated.
No candidate: `attachment_stripped` when a `/Keywords` string in clear
contains `faunosteo-archive`, else `no_attachment`. A stamped file saved
with object streams (before BLK-009) hides its stamp and reads as
+8
View File
@@ -88,6 +88,14 @@ hides its `/Keywords`, so a stripped attachment on such a file reads as
- Option A merged and the CLI on the scanner; no `PDFDocument.load` on any
untrusted input in the browser.
- A real-browser test of the inflate bound (Chromium, Firefox, Safari): the
vitest suite runs in Node and cannot see a browser inflater's read-ahead.
The audit of 60ee712 measured +2.2 GB RSS in Chromium on a 4 MB / 4 GiB
bomb before the inflater was fed in 16 KiB slices (+215 MB after); that
measurement is manual until a browser runner exists.
- Done: the candidate CPU/byte budget on the scanner, one 16 MiB budget of
examined content shared by every candidate, a monotonic `endstream` search
bounded to the cap, and no JSON parse for a candidate not opening on `{`.
- Option C or a server-side import (Step 14) with a hard memory and time
budget per file, for any path that still parses PDF objects.
- Dependency decision on pdf-lib (unmaintained since 2021): keep it for