Third security review LOW: the f32 spread of an arc is sampled only
when an outline is drawn and under its vertex cap; the variants'
extremes still join the bounds.
Second security review: endpoints equal in f32 draw no arc, as in
browsers; an arc carries the spread of its f32 variants (endpoint or
radius moved one ulp) in its deviation and their extremes in its
bounds. Probe and mutants extended.
Security review of the interior rule: an arc radius below f32
precision was scaled by the core to a half disc the browser never
fills. Nonzero radii below 2^-20 make the path unreadable; arcs add
their f32 drift to the chord deviation; tracing stops past the vertex
cap; a path without outline is refused.
A polygon or path zone or trigger must fill a disc of 4 CSS px on the
320 x 568 preview under the layer's nonzero rule, stroke left out:
conservative widest-disc search on the flattened outline, caps and a
per-sheet budget, closed codes interior_too_small and
marker_interior_too_small. Browser probe on six projects.
Fresh security review and verifier (MEDIUM): browsers read path data in
f32 and draw nothing for an arc radius or rotation past f32::MAX (WebKit
from 1e30) while the endpoints kept the verdict green. Every path and
viewBox number is now read within 2^20, else the path is unreadable.
Docs: section 12.12 transform table, mutation record, open zone finding.
Owner decision 2026-11-01: the scale bypass (viewBox blown up, drawing
scaled back) is closed by the visible size, 4 CSS px on the editor at
320 x 568 (equine junctions 4.97 px), verdict marker_too_small. A frame
is usable only where browsers draw it faithfully: edges within 2^20,
span from 1 unit, aspect 100, offset 256 spans; measured in 3 engines.
Owner decision 2026-11-01. The canvas paints and hits every point hitbox,
then every zone and trigger, then every visible twin, which captures on
what it paints; a junction's disc is no longer taken by the zone drawn
after it. The PDF view keeps the document order and its bytes. Click
spec: the five fixme run unmarked, six projects incl. a 320 px phone.
Owner-approved lot B (brief 2026-11-01). The composition root is the
engine's (namespaces and frame); the drawing's root becomes a group and
the marker layer and halo are its siblings, so no root property reaches
a marker. Point hitboxes declare stroke none and visibleFill. Census over
the reader's whole allowlist; R3 references compared under the declared
transform; click proofs in three engines.
Security review: the edge slack grew with the frame's offset (a frame at
1e15 accepted a circle 4000 units past its edge) and f64 min/max drop a
NaN bound. The slack now follows the span only; a non-finite bound is
unreadable, never in frame.
A view without view_box was framed by the drawing's viewBox: moving it and
translating the drawing back left the horse identical and its 19 main-view
markers out of frame, accepted. One frame definition (view_box, else the
drawing viewBox read with the SVG grammar) now frames both composers and is
measured by the verdict (verification.validation.out_of_frame) and the
preview; curves and arcs on their shape. The document is never rewritten.
Both composers append the marker layer as the root's last child, so a
root visibility="hidden" whose children set visible back drew the
drawing without a marker (0 marker pixels in Chromium and Firefox, PDF
and canvas). The reader refuses any root visibility but visible
(root_visibility); cleaning, verdict, preview, PDF gate and both
composers answer it.
A drawing defining filter#osteo-halo replaced the PDF halo (a transparent
filter erased the marker); an element holding the id before the first
defs won its resolution. The reader refuses the id in any letter case
(reserved_id), both composers refuse such a tree, the halo is always the
engine's. Exact name only: osteo-overlays stays.
The reader removes any attribute holding a url( that url_openings does not read, so reader, verdict, preview, canvas and PDF view accept the same drawings. maskUnits, maskContentUnits, clipPathUnits allowed with their two keywords (owner go). The owner's two raw Figma exports committed as fixtures with fixed expectations.
Composer check counts url( openings apart from the reader; tab and
newline character references tested; RUST.md sizes and counts of the
final build; owner procedure for Illustrator, Figma and Affinity.
Security review BLOCK(1): url(#%61) passed the dangling check and the
per-view scoping and resolved in browsers. One strict reading now, other
spellings are active content; glued names are other functions; the
composer refuses an unscoped reference; owner export procedure.
sheet.edit remove answers has_dependents, the document unchanged, while
dependents lists anything but descendants (owner decision 2026-10-29);
the property test checks no removal adds a referential issue.
Refuse local references to ids a drawing does not define, scope canvas
ids per view, upload every drawing in one stored form checked by the
verdict and the preview, keep strict numeric rgb() colours, and commit
real Inkscape 1.4.3 exports with their generator.
Verifier ECARTS(3) and security PASS on 8294d54: one indexed walk per
view chain in sheet.edit, a set for referenced drawings, released
preview parses, xml:base and CSS loaders refused, fail-closed PI scan,
PI drawings rejected by the preview, the builder naming its cap.
Security audit (PASS, 2 MEDIUM, 2 LOW) and verifier (ECARTS 6) on
741c25c: processing instructions and HTML fetchers refused, active
content under a long title classified, linear review groups, full
fixpoint check, bounded preview drawings, missing cap fixtures and
property checks, record wordings.
The sanitiser walk names removals that carried active or external
content; clean_svg answers clean, cleanable or rejected, the cleaned
text reading back with no removal. Committed cairo and ImageMagick
exports; the owner's editor exports go under exports/owner.
The core's authoring surface of contract §16.3: the server's document
measure, review groups with digests and removal groups, the change list,
the pure server verdict in the shape M5 reads, PreviewSheet with draft
and published modes, and the editor's commands with the §6.6 markings.
The TS pipeline, domain, upload script and legacy scripts leave; xmldom out of the
lockfile, Dockerfile copies trimmed. LEDGER.tsv (407 cases) and FILES.tsv (112 files)
record each disposition; r4 and r5 census tests guard the tree, lockfile and
production imports.
The Rust pipeline is the official seed writer (pnpm import:equine); sheet-ref.json
and report-migration-map.json carry 12d78872...74e2 and the v1 block, the only two
seed files changed. 27c7383c...2660 stays a registered alias, honoured only through
LoadedSheet::recognises_historical_stamp (verified v1, recomputed document). M1: the
snapshot of an old report keeps its stamp; export, archive reader strict. Historical
oracles run in their pre-switch context; equivalence is crate-private.
v1 refuses a non-hex or historical value (NotCanonical); the switch patch now builds
report-compare and refuses unknown block keys; RUST.md 11.19 states the measured
rehearsal (4 + 3 expected failures) instead of a green claim.
canonical_json (RFC 8785), sheet_fingerprint behind feature fingerprint,
TRUSTED_ALIASES (one equine pair), equivalence and check_declared; tests against
the RFC vectors, the V8 numbers, an independent Node reference and the archives.
No consumer wired: the switch is a rehearsed patch under .claude/tasks/plans.
check_v3_shape at read_report_document v3 and parse_envelope (sheet, practitioner
as non-objects, severity as non-string); Protocol errors final; DEV address must be
an own unicast address. SQL confirm function and DEV keys characterised, no change.
Core rules curation::upload (manifest, local destination, bytes through the SVG reader,
plan from observed state); binary upload-sheet-assets over ureq (handshake, no overwrite,
confirm after verify, bounded retries, key never printed); pnpm assets:upload repointed,
e2e:sheet-assets for the local stack. 4 MiB cap on changes.json and references.json.
FLOOR clarification for audit prose with its negative proof.