export_pdf_for_test on the proofs build (ExportInput JSON + base64 rasters): the 14 O2
runs give the native digests through wasm; the frozen O11 scanner reads every R4 PDF and
the N-1 Zod readers accept its envelope; a stripped R4 PDF answers attachment_stripped;
the TypeScript engine reads the R4 PDFs back to the R3 outcome; E2 size ≤ R3 + 5 %. Page
resources list the images the page draws (listing every embed cost 9 MB on CAP). Node,
stub rasters, TS → Rust: demo 198 → 36 ms, L 772 → 57 ms, CAP 5.7 s → 0.22 s; files
236 KB, 386 KB, 1.6 MB against 243 KB, 463 KB, 2.2 MB. Module with the writer linked
(proofs build): 531.6 KB gzip; shipped unchanged at 426 KB until R4.6 links it.
objects.rs (own minimal writer: numbered objects, streams, classic xref, pdf-lib's string
forms; pdf-writer not used: every number spelt through numberToString, no f32), serialize.rs
(catalog with names tree and AF, pages, 3 Type1 WinAnsi fonts, image XObjects with the
DeviceGray SMask, Link annotations with hex ASCII URIs, Info with Producer/Creator
Faunosteo, the Flate attachment with Params and Filespec), export.rs (checks → envelope
as JSON.stringify(_, null, 2) → layout → file → self-check through the R4 reader;
export_from_envelope). cargo pdf_serialize against the 14 R3 PDFs: content streams equal
after resource-name resolution, image planes and SMasks equal, links equal, Info equal
but Producer/Creator, attachment bytes equal (store path: values equal, D-R11 order);
qpdf --check clean, pdfdetach lists the attachment, poppler 150 dpi 0 pixel differs on
14 runs; R4 PDF → Rust import = R3 PDF import, re-export from the envelope byte-identical;
R4 output digests frozen in goldens/r4/pdf.*.sha256.
The branding gate in pngIssue's order; the decoder reproduces UPNG's decode and toRGBA8
(typed-array semantics, first-row filter rewrite, the Average first-row pre-pass at
absolute offsets, the 16-bit grey tRNS key at off + x, 16-bit high byte, palette tRNS by
declared length), pdf-lib's alpha drop; caps checked before any allocation; crafted
files refused (animated, text chunks without NUL, broken streams) where UPNG spun or
threw. 108 O6 rows equal; 50 hostile files return in under 20 ms.
drawText/drawImage/drawLine/drawRectangle as pdf-lib 1.17.1 emits them (cleanText, 24 TL,
upper-case hex, the doubled m of drawLine, the h f of a filled rectangle); numberToString
with the e-N / e+N branches and V8's Math.pow(10, k) one-ULP quirk at k = 23 and 210;
WinAnsi text encoding. The 12 R3 PDFs' 60 content streams (224,534 bytes) equal the
emitter's output from the O1 logs with only the resource names renamed; 703 O4 cases.
pdf::write: text (safe_text with the frozen V8 classes, js_upper from the O4 tables,
js_trim, measure = Σ(WX || 250 + KPX) × size / 1000 over WinAnsi bytes, wrap_lines on JS
\s), afm_data generated by scripts/r4-afm-data.mjs from O5 and O4, dates (Date.UTC years
0-99 → 19xx, fr-FR / en-US long dates), input checks (branding and labels schemas,
snapshot gate with the R3 SVG reader, sheet mismatch, page caps), the header, footer,
observation and conclusion layouts into a DrawLog (per page, footers last, links by page).
cargo: O4 55,992 widths bit-exact, 1,168 wraps, safeText over every code point, upper
1,580, O6 6,468 cells + 22; O1 12 runs 2,124 ops equal at full precision; the Step 6 text
and image goldens (24) derive from the log. ts-rs: ExportInput, ExportErrorKind,
ExportWarning, LogoSlot.
docs/development/R4_J1.md: the two LAN surfaces (dev server, production build behind the
DEPLOY CSP through serve.mjs --host on the DEV origin), the kit folder generated by
scripts/r4-j1-kit.mjs (reference PDFs, the O8 hostile corpus with expected messages,
rescue copies), twelve checks per device, what to report.
Same perimeter as the TS baseline (226 files, 3 engines, real Home page): 666/678 outcome
and message agreements, the 12 others the 4 declared files; W5 0 long tasks (TS 91 and
173 ms); 50 MiB file in WebKit 145/126 MiB against 324 with the worker read and 76 for the
TS. CSP: editor opens with the keyword, the engine message without, in 3 engines, the
worker's wasm-eval refusal reported by Chromium and Firefox. Leak series on the app:
Chromium ≤ 2 MiB per import after init, Firefox GC sawtooth reclaimed, WebKit grows with
same-tab navigations (control too) and levels off on one page; 20/20 WebKit workers closed.
Page messages validated by field (size and chunk length as safe byte counts, length within
the buffer); the task post wrapped so a refused clone leaves no orphan worker; a second
ready ignored; the 16 MiB copy cap enforced in facade::restore_copy as well as in the
worker (the uncommitted part of R4.3a). Gate verdict PASS, 7 LOW.
Measured at R4.3a (RUST.md §10.5): reading a file inside a WebKit worker costs about
180 MiB of content RSS whatever the method (arrayBuffer, stream, BYOB), 14 MiB on the
page; a page-side BYOB read feeding the worker's FileBytes gives 106 MiB on a 50 MiB file
against 272 in the worker and 76 for the TypeScript. Protocol v1 gains chunk / ack / end:
one 1 MiB buffer travels (transferred) and comes back after each push, so the page holds
one chunk and the worker's sink the only whole copy; no byte is parsed on the page. The
worker refuses an oversize announced before any chunk and a size that changes while
reading. Tests adapted (handler, client with a scripted ack, read failure on the page).
Streaming into FileBytes removes the file's JavaScript copy in Chromium and Firefox
(50 MiB: 65 and 67 MiB summed growth against 56 and 79 for the TypeScript); WebKit keeps
about 180 MiB for any file read inside a worker (272 against 76), a main-thread BYOB read
feeding the worker's sink measured at 106. Corpus 666/678 on the stream path. Leak series
(3 engines × 3 documents × 3 repetitions): Chromium ≤ 0.5 MiB over the last 5, Firefox
within ±3 (idle only), WebKit copy +13 then level (12-import series −2.3). 20 dense
imports in WebKit: no worker error event.
Protocol v1 (task → ready → result | failed with phase and kind), the worker entry with
the glue bundled statically and the page's compiled Module posted, the file streamed into
the module's memory; client with start and task timers (paused while hidden), settle-once,
terminate always, error events never masked; module-level controller (one import at a
time, cancel on leaving Home, worker chunk prefetched once the engine is loaded); the
settlement mapped to the B.1.8 messages (OD-5: a trap, a timeout or an unreadable file is
« Erreur de lecture du PDF. », never a corrupt PDF). HomeRoute keeps its routing and size
pre-checks, loads its own index after acceptance, drops a result when the session changed.
Tests: handler and client in-process, the corpus message test (O8 51, O7 stored inputs,
O2/O3/Step 5), the six existing suites migrated to the real worker logic.
FileBytes gathers a file chunk by chunk in linear memory (capacity reserved from the
announced size, refused above the PDF cap); SheetHandle.import_pdf_bytes and
restore_copy_bytes consume it, so the worker streams file.stream() and the file lives
once, never as a JavaScript ArrayBuffer nor as a JSON-wrapped string. restore_copy:
strict UTF-8, one BOM, the OD-3 budget, draft.restore's kinds. Node: same envelopes as
import_pdf on the 51 O8 cases; rescue-copy parity with draft.restore.
r4-import.mjs runs the 226 baseline files and the hostile classes through a disposable
module worker calling SheetHandle.import_pdf in Chromium 156, Firefox 157 and WebKit 27.2
under the RSS watchdog (not a guaranteed limit); compare-imports.mjs prints TS vs Rust;
serve.mjs gains opt-in mounts (/r4/, /pkg/). Results 2026-10-18-r4-import.json: 666/678
agreements, the 12 others are the 4 declared files; W4 met (max 68 MiB linear); the
1 GiB ceiling traps in every engine.
Security gate finding (R4.2): prescan stopped counting once the depth exceeded 64, so a
16 MiB rescue copy nested 65 deep passed the value budget with 66 values and built its
792 MiB tree. The count now runs to the end (the depth verdict unchanged, the PDF probe
unchanged); the class is in the budget suite (48 MiB, refused) and pinned in cargo.
69 ported rows of scan, inflate, read and the security import cases now name the O7/O9
cargo tests; r4-ledger.mjs refuses a ported cargo target absent from cargo test --list.
json_js::MAX_JSON_VALUES from the R4.2e measurements (CAP envelope 39,652 values, 612-leaf
export 23,803; ~100 B of tree per value, a 16 MiB dense document 784 MiB with the cap off,
16 MiB with it); draft.restore refuses an over-budget copy as invalid; --max-memory 1 GiB
proven to trap on the proofs build; cargo pdf_budget prints the value table, the Node
budget suite prints the per-class cost (fresh instance, W4 asserted).
locate_attachment (shared 16 MiB budget, JS-path steps in the TypeScript order, OD-3 cap
hook off until measured), pdf::import, facade import_pdf with the PdfImportKind kinds
(ts-rs exported). cargo: O7 144 inputs (locate, read, archive rows) and O8 51 cases equal
to the R3 path, 3 OD-4 differences declared (lone surrogate in an overwritten duplicate
value). Node through the shipped module: the same O8 outcomes, every reference PDF.
Module 421.7 KB gzip (W1 budget 600 KB).