fix(anatomy): R5.4 aligned report tables bounded before they are built, declared case fully held, docs

This commit is contained in:
bchanot committed 2026-10-23 14:39:51 +02:00
1 parent aeb5050502
commit 777c8000b8
8 files changed
+215 -21

No files matched your search

@@ -0,0 +1,153 @@
# Security gate R5.4, round 6 (fresh, blind)
Branch `feature/rust-r5-pipeline-core`, HEAD `aeb5050`. Scope:
`git diff c6e0d85..HEAD -- crates/core packages/anatomy/tests docs/development/RUST.md`
(25 files). Mode: audit (this file is the only write). The brief carried a list of earlier
findings; per the auditor contract it was not used as a checklist. The scan below is complete
on its own and covers those areas anyway.
```
SECURITY — VERDICT: PASS
TOOL: semgrep 1.168.0 — p/security-audit, p/secrets, p/owasp-top-ten
SCOPE: 25 files
BLOCKING:
(none)
REPORTED (non-blocking):
- [MEDIUM] resource/changelog-table-amplification — crates/core/src/curation/draft/changelog.rs:100 (via crates/core/src/curation/markdown.rs:133)
- [LOW] resource/per-view-drawing-amplification — crates/core/src/curation/draft/pipeline.rs:204
- [LOW] resource/guard-underscore-pairing-quadratic-in-isolation — crates/core/src/curation/report_guard.rs:216
- [LOW] markdown/invisible-hangul-fillers-admitted — crates/core/src/curation/report_guard.rs:159
- [LOW] markdown/unpadded-wiki-link-in-table-cell (context, by design) — crates/core/src/curation/report_guard.rs:58
- [LOW] anti-gaming/clippy-allow-in-new-test-file (context, convention) — crates/core/tests/curation_markdown.rs:6
- [LOW] anti-gaming/clippy-allow-in-new-test-file (context, convention) — crates/core/tests/curation_review.rs:6
PROOF: semgrep 162 rules on 25 files → 0 findings; checklist 13 checks → 7 findings
```
## What was run
- semgrep 1.168.0, the three pinned packs, `--metrics=off`, on the 25 existing files of the
diff: 162 rules, 25 files scanned, 0 results, 0 errors.
- Checklist greps on the added lines: secrets, SQL concatenation, unsafe render, logging,
suppressions, tracked `.env` / credential files (only `.env.example` is tracked). Nothing
found. No endpoint in scope (authz check N/A).
- Guardrails scanner (`--domain api,supabase`): 0 FAIL, 1 WARN, 1 MANUAL. The WARN
(`SB-POLICY-NOT-OWNER`, `supabase/migrations/20261008110000_anatomical_sheets.sql:1177`) is
outside the diff, not counted.
- `cargo test -p faunosteo-core --features curation`: every suite green. `cargo clippy
--all-targets -D warnings` (curation): clean. No `std::fs`, `std::env`, clock or process use
in the new core modules (only the binary, which already carried its documented allow).
- `import-equine --check`: legacy 6 files 0 differences; draft 12 files 0 differences, one
`declared:` line for `review/app-main.svg`.
- Differential fuzz of `report_guard::format_report` against Prettier 3.8.3 (the TypeScript
formatter, same options as `packages/anatomy/src/format.ts`), through a scratch binary on the
core library. 10,201 generated reports (seeded random strings over the admitted alphabet plus
Markdown tokens: URL starts, entities, `**`, `[[`, `](`, `<a`, `{%`, `[^1]`, `:warning:`,
line-start markers), placed in the list-item fields (with the pipeline's own
`assert_single_line` / `assert_authored` pre-filters) and in CHANGELOG table cells (with the
`cell()` escape). 2,226 admitted, 7,975 refused. 0 admitted report differs from Prettier's
output. AST census of every admitted report (Prettier's own Markdown parser): heading, list,
item, table, row and code-span counts always equal the template's; no link other than a bare
URL autolink literal, no image, HTML, emphasis, footnote, definition, break, liquid node.
One `wikiLink` node, in a table cell (see LOW 5).
- Table width parity: 28 unusual letters (combining marks with the Alphabetic property, Indic
vowel signs, CJK, fullwidth, halfwidth, Hangul jamo and fillers, non-BMP letters) in a table
cell: 26 admitted and identical to Prettier, 2 refused. No silent alignment difference.
- Guard timing on 12 pathological 0.4 to 1 MB lines (debug build): 32 to 224 ms each, linear.
- `--out` / `--check` file safety, 18 probes in the scratchpad: fresh folder (18 files, `diff
-r` identical but the declared `app-main.svg`); non-empty folder, symbolic link (plain,
trailing `/`, trailing `/.`), dangling link (target not created), path inside the seed,
inside the seed through a linked parent, link to `draft/`, `..`, empty path, existing file,
missing parent, `/`, extra arguments: every one refused, nothing written; non-absolute
`INIT_CWD` falls back to the current folder. The seed and the git tree are unchanged.
- Resource probe on the real equine pipeline (scratch binary, release build, `ulimit -v`).
## Findings
### MEDIUM 1: the CHANGELOG grows as rows × the widest cell of each column
`changelog.rs:100` `level_section` builds one table per level, and `format_report` aligns it
(`markdown.rs:133` `align`, `:163` `padded`): every row is padded to the widest cell of its
column. Titles and rationales are authored and uncut (only `Avant` / `Après` are cut to 120
units). One change with a long rationale plus another change producing many records makes every
row as wide as that rationale. Measured on the equine pipeline, all level A:
| added to changes.json | records | CHANGELOG | max RSS |
|---|---|---|---|
| +25 KB (200 subtree sets) | 5,226 | 7.8 MB | 45 MB |
| +45 KB (same, one 20 KB rationale) | 5,226 | 98 MB | 233 MB |
| +105 KB (same, one 80 KB rationale) | 5,226 | 381 MB | 788 MB |
| +155 KB (600 sets, one 80 KB rationale) | 14,026 | 1.09 GB | 2.2 GB |
A rationale cloned into many records (the R5.3 shape) aborted at 4.4 GB under a 6 GB
`ulimit` from +175 KB. Within the 16 MiB read cap the demand reaches terabytes: certain OOM of
the tool, and `--out` would try to write a file of that size (into a fresh folder, never
overwriting). Prettier aligns tables the same way, so the TypeScript had the same output. This
is not the declared R5.3 residual (records cost ops × items × value size): here a single wide
cell multiplies every row, with the records themselves staying small. No integrity or
confidentiality impact; a local developer tool fed by an authored file that goes through review.
Hint: compute the aligned size (rows × the sum of column widths) before building the strings and
refuse past a stated cap with a loud `PipelineError`, or cap authored title / rationale length.
Declare the bound in RUST.md next to the 200,000-value one.
### LOW 2: two drawings per view, no view count bound
`pipeline.rs:204` `review_drawings` writes an overlay and an application drawing per view, each
a clone of the whole cleaned drawing (`overlay.rs:350`, `app_review.rs:27`), about 137 KB each,
all held in memory before any write or comparison. A detail view needs about 25 document values
(view, view box, trigger annotation), so the 200,000-value bound allows about 7,700 views:
around 2.1 GB of output. Estimated, not measured. Same output as the TypeScript. Hint: the same
output-size cap as MEDIUM 1 covers it.
### LOW 3: `underscore_pairs_later` is O(parts × length) on its own
`report_guard.rs:216`: for each plain part holding `_` it rescans every later code span. 40,000
`a_b \`cN\`` parts (509 KB) took 213 s (debug). The generator never writes such a line: at most
two authored parts per line (title + rationale, question) and the separators between code spans
hold no `_`; the realistic worst case, one underscore part followed by 100,000 code spans (1 MB),
takes 224 ms. So the "linear guard" holds for the generator's line shapes only. Hint: one
backward pass that records whether a later code span can close.
### LOW 4: invisible Hangul fillers pass the allowlist
`report_guard.rs:159` admits `char::is_alphanumeric`, which includes U+3164, U+115F, U+1160 and
U+FFA0 (Hangul fillers, category Lo, rendered blank). An authored title or rationale can carry
invisible runs in REVIEW-REPORT.md. No structure is forged and Prettier keeps them (identical
output). Hint: refuse Default_Ignorable_Code_Point letters with the other refused classes.
### LOW 5 (context, by design): unpadded `[[x]]` admitted in a table cell
Only padded wiki links are refused in cells (`report_guard.rs:58`); list items refuse `[[`
and `]]` whole (`review_report.rs:364`). Prettier parses `[[x]]` in a CHANGELOG cell as a
`wikiLink` and prints it back byte for byte, so the output matches the TypeScript. Recorded so
the per-cell rule stays a conscious choice.
### LOW 6 (context): clippy allow attributes in the two new test files
`tests/curation_markdown.rs:6` and `tests/curation_review.rs:6` add
`#![allow(clippy::disallowed_methods, clippy::unwrap_used, clippy::expect_used)]`. They are not
security-scanner suppressions (no `nosemgrep`, `nosec`, `eslint-disable`): they are the
convention `clippy.toml` documents ("Tests and the worker binary allow these lints locally where
they own I/O"), present in the same form in all 34 existing integration test files. Not
blocking.
## Areas checked and found sound
- Report structure: no authored text reaches a line start (control characters refused in
list-item fields, `\n` → space and `\r` refused in cells); code spans come from the generator
only (authored backticks refused, content limited to `[A-Za-z0-9_:/. -]`, field segments to
`[A-Za-z0-9_-]`, change ids `CHG-\d{3,}`, reference ids declared and stable, ids written by an
op checked at the op); `](`, `][`, `![`, `<` before a letter / `/` / `!` / `?`, character
references, `{%`, `%}` refused; `*` only as `**word**`; `_` only intraword plus the
code-span closer rule; backslash only before `"` or `|`; table cell count checked.
- SVG: elements and attribute names are fixed or come from the frozen drawing read by the
core's reader (refused if anything is removed); values escaped (`& < > "`, `\t \n \r` as
references in attributes); XML-invalid code points and CR refused (`drawing.rs:41`); file
names built from stable view ids.
- `--check`: the tree fallback only for `review/app-*`, only with the declared rule, and only
when the committed file rebuilds from its tree byte for byte, the rule changes it and the
result equals the produced bytes; the committed file goes through the core reader's prescan
first.
- `--out`: `..` refused, final symbolic link refused (trailing slash handled), parent
canonicalised, seed refused by path and by (device, inode) of every seed folder, existing
destination identity-checked and must be empty, folders `create_dir`, files `create_new`,
components all `Normal`. The swap window between check and use stays as declared (§11.13).
@@ -49,6 +49,12 @@ Fresh verifier and security on the fixed commit: below.
- Security (`.claude/audits/2026-10-23-rust-r5-review-security-5.md`): MEDIUM F2 the same wiki link across fields; MEDIUM F1 an authored field segment (`name.x]`) inside a generated code span closing an author's `[`; LOW docs "code spans are the generator's alone" wrong (segments are authored); LOW quadratic `in_url_token` and padded-wiki scans; LOW test allows (context).
- Fixed: field segments limited to `[A-Za-z0-9_-]` at the op (declared refusal; TypeScript accepted any); code spans admit only the generator alphabet `[A-Za-z0-9_:/. -]`; list-item fields refuse `[[`, `]]`, `{%`, `%}` whole; plain parts refuse `{%`/`%}`; URL-token membership in one pass, padded wiki through one find/rfind; sweep 798 probes (eight cross-field construct pairs): 268 identical, 530 refused, 0 silent; segment refusal test with its reason; docs counts (draft 4,385, gain 1,477, report_guard.rs 258) and rules restated.
## Round 6
## Round 6 on aeb5050: verifier C1 to C8 MET (C9 unverifiable: gates run in parallel; 8 FLOOR lines = keywords quoted in audit prose), security PASS (1 MEDIUM, 6 LOW); `pnpm check:all` green 14:02:00
- Verifier: 0 silent over 100,000 guard probes × 3 contexts, 45,000 cross-field pairs (28 guard-level, all stopped first by `assert_single_line`), 32,000 end-to-end changesets TS vs Rust, 148,921 letter/digit code points in a cell. Observations: stale `draft/mod.rs` header; RUST.md: `format_markdown` gained the cell-count refusal (import report too), "combining marks" wording (alphabetic marks pass, aligned alike); declared emoji case skipped its review report and drawings; TODO R5.4a cites `format_markdown`.
- Security (`.claude/audits/2026-10-23-rust-r5-review-security-6.md`): MEDIUM CHANGELOG table amplification (rows × widest cell: +155 KB changeset → 1.09 GB, 2.2 GB RSS; TypeScript alike); LOW per-view drawings bounded only by the value bound (≈2 GB estimate); LOW `underscore_pairs_later` O(parts × length) in isolation, unreachable (≤2 authored parts per line); LOW invisible Hangul fillers admitted (Prettier keeps them); LOW unpadded `[[x]]` in a cell by design; LOW test allows (convention).
- Fixed: aligned table bounded at 32 Mi display columns before any row is built (`Guard`, test `an_aligned_table_past_the_bound_is_refused_before_it_is_built`; equine widest ≈0.5 MB), declared in §11.14; residual risks (drawings bound, Hangul fillers) in §11.15; `draft/mod.rs` header; RUST.md wording (format_markdown shared refusals, alphabetic marks); declared emoji case now holds its review report (same text or declared refusal) and drawings; TODO R5.4a.
## Round 7
Fresh verifier and security on the fixed commit: below.
+1 -1
View File
@@ -321,7 +321,7 @@ feat/findings-tree-v2 merged (FindingsTreePicker tracked in HEAD); all work belo
- [ ] OWNER DECISION (found 2026-10-23 by the R5.3 final verifier): serde derives of the R2 contracts read a struct from an array and a unit enum from a one-key map, where Zod refused; likely reachable by the gated readers of imported PDFs (`archive/read.rs`, `report/read.rs`, `pdf/import.rs`); probe the readers, then a per-file go for a strict deserialisation (RUST.md §11.13)
- [ ] chore (pre-existing, security gates R5.3): `clippy.toml` disallowed list misses `fs::create_dir*`, `OpenOptions::open`, `read_dir`, `canonicalize`, `remove_*`, `env::current_dir`, `env::args_os`, `process::Command`; extend so the core's no-I/O rule is enforced, not only observed
- [ ] R5.4 review overlays + app drawings, CHANGELOG + REVIEW-REPORT, --check over the whole draft (owner go 2026-10-23, conditional on the R5.3 final checks: met on c6e0d85; contract `.claude/tasks/contracts/2026-10-23-rust-r5-review.md`; stop at the checkpoint, no R5.5)
- [x] R5.4a `draft/{labels,changelog,review_report}.rs` + `stats::stat_rows`: CHANGELOG and REVIEW-REPORT raw text = ledger pre-Prettier digests, formatted text byte-exact (`markdown::format_markdown`); compactValue on UTF-16 units
- [x] R5.4a `draft/{labels,changelog,review_report}.rs` + `stats::stat_rows`: CHANGELOG and REVIEW-REPORT raw text = ledger pre-Prettier digests, formatted text byte-exact (`report_guard::format_report`: allowlist guard + `markdown::format_markdown` alignment); compactValue on UTF-16 units
- [x] R5.4b `draft/{overlay,app_review}.rs` on the core tree (`read_svg`, frozen horse.svg, prolog kept verbatim): review overlays and app drawings; bytes vs xmldom where reproducible, parsed trees otherwise; the R3 trigger-name decision declared for app-main
- [x] R5.4c `render_draft_artifacts` complete (TS order); `import-equine --check` covers the 12 draft files with explicit criteria (bytes; declared tree rule), PENDING list and `shortcut:` removed; fails on any missing or unauthorised difference
- [x] R5.4d proofs: `tests/curation_review.rs` (the 6 files, ledger digests, determinism); TS generator `r5.review-synthetic-oracles.test.ts` → `oracles/r5/review-synthetic.json` (changelog raw + formatted, review report, overlays, app drawings on mini-sheet cases: Unicode, pipes, newlines, long values, emoji refused, every status); R5.3 LOW carry-over (destination identity, cite membership, unreadable-seed message, stale comment, doc nits)
+5 -5
View File
@@ -1,8 +1,8 @@
//! The draft side of the pipeline (R5.3): the authored changeset applied on the legacy
//! baseline, the report migration map, the document statistics and the seed SQL
//! (`packages/anatomy/src/curation`, without the review drawings and the Markdown reports of
//! R5.4). Parity with the TypeScript is proven by `tests/curation_draft.rs` against the
//! committed `draft/` folder, `oracles/r5/pipeline.json` and `oracles/r5/curation-synthetic.json`.
//! The draft side of the pipeline (`packages/anatomy/src/curation`): the authored changeset
//! applied on the legacy baseline, the report migration map, the document statistics and the
//! seed SQL (R5.3), the CHANGELOG, the review report and the review drawings (R5.4). Parity with
//! the TypeScript is proven by `tests/curation_draft.rs` and `tests/curation_review.rs` against
//! the committed `draft/` folder and the `oracles/r5` files.
pub mod app_review;
pub mod apply;
+13
View File
@@ -130,6 +130,12 @@ pub(crate) fn split_row(line: &str) -> Vec<String> {
.collect()
}
/// The largest aligned table written, in display columns (rows × padded row width): every row
/// is padded to the widest cell of each column, so one long authored cell multiplies by the row
/// count (an 80 KB rationale over 14,000 records asked for 1 GB). The equine CHANGELOG's widest
/// table holds about 0.5 MB.
pub const MAX_TABLE_COLUMNS_TOTAL: usize = 32 * 1024 * 1024;
fn align(lines: &[&str]) -> Result<Vec<String>, PipelineError> {
let rows: Vec<Vec<String>> = lines.iter().map(|line| split_row(line)).collect();
let columns = rows.first().map_or(0, Vec::len);
@@ -144,6 +150,13 @@ fn align(lines: &[&str]) -> Result<Vec<String>, PipelineError> {
widths[column] = widths[column].max(display_width(text)?);
}
}
let row_width = widths.iter().sum::<usize>() + 3 * columns + 1;
if row_width.saturating_mul(rows.len()) > MAX_TABLE_COLUMNS_TOTAL {
return Err(PipelineError::Guard(format!(
"an aligned table of {} rows × {row_width} columns exceeds the {MAX_TABLE_COLUMNS_TOTAL} bound",
rows.len()
)));
}
let mut out = Vec::with_capacity(rows.len());
for (index, row) in rows.iter().enumerate() {
out.push(if index == 1 {
+14
View File
@@ -77,6 +77,20 @@ fn tables_align_as_prettier_aligns_them() {
assert_eq!(refused, vec![5], "only the emoji table is refused");
}
/// One long cell pads every row of its column: the aligned table is bounded before any row is
/// built (security gate R5.4 round 6: an 80 KB rationale over 14,000 records asked for 1 GB).
#[test]
fn an_aligned_table_past_the_bound_is_refused_before_it_is_built() {
let mut raw = vec!["| a | b |".to_owned(), "| --- | --- |".to_owned()];
raw.push(format!("| {} | x |", "w".repeat(2_000)));
raw.extend((0..20_000).map(|_| "| a | b |".to_owned()));
let error = format_markdown(&raw.join("\n")).unwrap_err();
assert!(
matches!(&error, PipelineError::Guard(message) if message.contains("exceeds the")),
"{error}"
);
}
#[test]
fn slugs_and_uuids_match_the_typescript() {
let semantics = oracle("js-semantics.json");
+14 -11
View File
@@ -361,18 +361,21 @@ fn the_synthetic_review_cases_match_the_typescript_or_are_refused_as_declared()
),
"{name}"
);
continue;
} else {
assert_eq!(
rendered.changelog.unwrap(),
case["changelog"].as_str().unwrap(),
"{name}"
);
}
// The review report and the drawings of a declared case are held too: the same text
// or a declared refusal, never another text.
match rendered.review {
Ok(review) => assert_eq!(review, case["review"].as_str().unwrap(), "{name}"),
Err(PipelineError::UnsupportedWidth(_) | PipelineError::UnsupportedMarkdown(_))
if DECLARED_REFUSALS.contains(&name) => {}
Err(other) => panic!("{name}: {other}"),
}
assert_eq!(
rendered.changelog.unwrap(),
case["changelog"].as_str().unwrap(),
"{name}"
);
assert_eq!(
rendered.review.unwrap(),
case["review"].as_str().unwrap(),
"{name}"
);
let views = typed(&case_document(&frozen, case)).views.len();
assert_eq!(rendered.drawings.len(), 2 * views, "{name}");
let rule = TriggerNames::of(&typed(&case_document(&frozen, case)));
+8 -3
View File
@@ -3156,7 +3156,8 @@ three security probes, each round finding a Prettier rewrite the previous rules
also rewrites Markdown in authored text (escapes, emphasis, whitespace, links); the port
reproduces only the table alignment. After a denylist that leaked round after round, the guard of
the draft's two reports (`curation/report_guard.rs`, `format_report`; the R5.2 import report keeps
`format_markdown`, verified at R5.2) is an allowlist: outside code spans, letters and digits,
`format_markdown`, verified at R5.2, which gained at R5.4 the cell-count refusal and the table
bound below, both shared with the import report) is an allowlist: outside code spans, letters and digits,
single spaces (none at a line end), the punctuation `.,;:!?'"«»()[]{}/-–—…%°≠=+#<>’‘“”±→×≤≥|`, an
`&` before a space or as a URL's query separator (`&name=` in a bare URL, the name not starting
with one of HTML's 106 legacy references, which Prettier decodes without a `;`), an underscore
@@ -3200,8 +3201,9 @@ pass the guard unchanged.
| Input | TypeScript | Rust |
|---|---|---|
| authored text outside the allowlist: Markdown Prettier rewrites (`*x*`, `_x`, `x_**a**`, `**a****b**`, `Voir_https://…`, `a b`, a non-ASCII space, a tab, a line break in a list item, `` `` ``, `a\|b`, `[a](<b>)`, `[[ a ]]`, `&#42;`…), the constructs refused whole (links, HTML tags, autolinks, character references) and any character outside the admitted set (combining marks, pictographs, symbols) | written, then rewritten by Prettier (escapes, `_emphasis_`, collapsed spaces, indented continuation lines, normalised links; a row with an extra cell), or kept | refused (`UnsupportedMarkdown`), even where Prettier would have kept the text |
| authored text outside the allowlist: Markdown Prettier rewrites (`*x*`, `_x`, `x_**a**`, `**a****b**`, `Voir_https://…`, `a b`, a non-ASCII space, a tab, a line break in a list item, `` `` ``, `a\|b`, `[a](<b>)`, `[[ a ]]`, `&#42;`…), the constructs refused whole (links, HTML tags, autolinks, character references) and any character outside the admitted set (pictographs, symbols, combining marks without the Alphabetic property; an alphabetic mark such as U+0345 is a letter and Prettier aligns it alike) | written, then rewritten by Prettier (escapes, `_emphasis_`, collapsed spaces, indented continuation lines, normalised links; a row with an extra cell), or kept | refused (`UnsupportedMarkdown`), even where Prettier would have kept the text |
| a pictograph in a table cell (an emoji, and ©, ®, ™, ✓, ★, ⚠ and the other characters `text::is_emoji_candidate` lists) | measured with Prettier's tables | refused (`UnsupportedWidth`, R5.2): the set is wider than the emoji, the refusal stays loud |
| an aligned report table past 32 Mi display columns in all (rows × padded row width; the equine CHANGELOG's widest table is about 0.5 MB): one long authored cell pads every row of its level, so an 80 KB rationale over 14,000 records asked for 1 GB (security gate R5.4 round 6) | Prettier aligned it (the same amplification) | refused before any row is built (`Guard`, test `an_aligned_table_past_the_bound_is_refused_before_it_is_built`), import report included |
| an id an `add` writes (the item's, and those nested in its `children` or `categories`) or a `rename` id that is not a stable id (a region code for a region), or a field segment outside `[A-Za-z0-9_-]` (a control character, NUL included, a backtick, a `]`, a space, a non-ASCII letter) | accepted when a later op removed or renamed the item: the raw id stayed in the records and reached the reports' code spans (a line break there forged a section, security gate R5.4 round 2) | refused at the op, the check the final document applies anyway (test `ids_and_field_segments_that_could_forge_report_structure_are_refused`) |
| a label, id or path holding a code point XML 1.0 refuses, or a carriage return, in a review drawing | written (a non-well-formed file; a raw CR read back as a line feed) | refused: the drawing is not written |
@@ -3238,7 +3240,10 @@ ruled and frozen in the sweep, so a combination of admitted characters no probe
differ; the comparison of the application's drawings accepts
exactly the trigger-name rule, byte for byte; the frozen
`horse.svg` is read, never regenerated, so a future change of the drawing needs its own
decision; the residual risks of §11.13 stand.
decision; the review drawings are bounded only through the document's 200,000 values (two drawings
of about 137 KB per view, about 2 GB at the bound, an estimate of the security gate R5.4 round 6,
not measured), and the allowlist admits the invisible Hangul fillers (U+3164, U+115F, U+1160,
U+FFA0, letters Prettier keeps, no structure); the residual risks of §11.13 stand.
Conditions for R5.5: the owner's validation of this checkpoint; the upload tool (`ureq`, the
service key from the environment, bucket `anatomical-sheet-assets`, key `<sha256>.svg`,