fix(anatomy): R5.4 records bounded in total, CHANGELOG tables projected before they are joined

This commit is contained in:
bchanot committed 2026-10-23 14:58:40 +02:00
1 parent 777c8000b8
commit 5880dfa959
9 files changed
+258 -6

No files matched your search

@@ -0,0 +1,88 @@
# Security gate R5.4 (round 7): review files in Rust
Branch `feature/rust-r5-pipeline-core`, HEAD `777c800`. Scope: `git diff c6e0d85..HEAD -- crates/core packages/anatomy/tests docs/development/RUST.md` (26 files). Blind, complete scan. No prior verdict was read.
```
SECURITY — VERDICT: BLOCK(2)
TOOL: semgrep 1.168.0 — p/security-audit, p/secrets, p/owasp-top-ten
SCOPE: 26 files
BLOCKING:
1. [HIGH] resource-bound/output-amplification — crates/core/src/curation/draft/changelog.rs:68 (record_row), :116-119 (table built whole); crates/core/src/curation/markdown.rs:52-63 (table), :139-159 (bound checked after the raw rows exist); crates/core/src/curation/draft/pipeline.rs:247-252; docs/development/RUST.md:3206 — the CHANGELOG raw table (rows × authored rationale/title) is fully built, joined and split before MAX_TABLE_COLUMNS_TOTAL is checked, so the declared case "80 KB rationale over 14,000 records" still exhausts memory: a 1.9 MB changes.json aborts at a 3 GB address-space cap; RUST.md's "refused before any row is built" is false — hint: bound before render_changelog builds a row (project rows × widest authored cell per level from the records, or cap the raw length while table() accumulates), and pin it with a pipeline-level test (build + render of the declared case under a memory ceiling), not format_markdown on a hand-built table
2. [HIGH] resource-bound/record-amplification — crates/core/src/curation/draft/ops.rs:197 (rename), :384 (remove); crates/core/src/curation/draft/records.rs:18-28 (out of diff, same pipeline) — every rename/remove record deep-copies the whole item (before and after) and every record clones the ChangeSummary (title, rationale, question); MAX_DOCUMENT_VALUES (200,000) bounds the document after an op, never the sum of the records: an 810 KB changes.json (one 5,000-child add, 400 renames back and forth) reaches 4.3 GB RSS at build, linear at ~10.7 MB per rename; ~1,400 renames exceed the 15 GB dev machine; R5.4's CHANGELOG then serialises each whole item again (compact_value) — hint: bound the records (total values or bytes recorded, checked as each record is pushed, refused with Guard) or record ids/paths and share the ChangeSummary (index or Rc) instead of cloning
REPORTED (non-blocking):
- [LOW] new lint relaxation `#![allow(clippy::disallowed_methods, clippy::unwrap_used, clippy::expect_used)]` — crates/core/tests/curation_markdown.rs:6, crates/core/tests/curation_review.rs:6 (integration tests only, same line as every existing test file; not a security-rule suppression)
- [LOW] pre-existing suppression `#![allow(clippy::disallowed_methods)]` — crates/core/src/bin/import_equine.rs:17 (the tool's declared I/O file, context)
- [LOW] error echo size — crates/core/src/curation/report_guard.rs:29-31 (UnsupportedMarkdown carries the whole refused line, multi-MB on stderr; authored content, no secret)
- [LOW] subfolder race — crates/core/src/bin/import_equine.rs:353-389 (write_exclusive re-checks the claimed root only, not the subfolders this run created; same-user swap, declared out of model in RUST.md §11.13; unchanged by this diff)
PROOF: semgrep 128 rules on 26 files → 0 findings; checklist 12 checks → 6 findings (2 HIGH, 4 LOW)
```
## Checklist detail
| Check | Result |
|---|---|
| Hardcoded secret, token, key, credentialed URL | none in the added lines |
| SQL by concatenation | none (no SQL in the diff) |
| Unsanitized render (innerHTML, eval, raw) | none |
| Sensitive endpoint without authz | not applicable (offline CLI and pure core) |
| Stack trace or internal path to a user | none (CLI messages to the operator) |
| Secret or PII in a log | none; `println!` lines print file names and counts |
| Tracked `.env` or credential file | only `.env.example` tracked |
| guardrails `scan.py --domain api,supabase` | 0 FAIL, 1 WARN (`SB-POLICY-NOT-OWNER`, `supabase/migrations/20261008110000_anatomical_sheets.sql:1177`, outside the diff: dropped) |
| Anti-gaming (new suppressions) | two test-file lint allows, convention of all test files, LOW |
| Report structure forgery (sections, list items, code spans, links, HTML) | held, see below |
| SVG well-formedness and injection | held: serializer escapes `& < > "` and CR/TAB/LF in attributes, `first_unwritable` refuses non-XML-1.0 code points and CR, element and attribute names are fixed, no href or event attribute written |
| Resource bounds and file safety | 2 HIGH (above); quadratic scans and `--out`/`--check` held, see below |
## Finding 1: CHANGELOG raw table amplified before its bound
`render_changelog` builds one row per record, each carrying the change's title and rationale (`changelog.rs:89`, `:94`). `table()` joins every row, wraps it, and joins again. `format_report` then guards each line, and `align` splits every row into owned strings. Only after all of that does it compare `row_width × rows` with `MAX_TABLE_COLUMNS_TOTAL`. So the bound caps the padded copy only. The raw table, which already holds N × R bytes, gets built four to five times first.
Measured with the full pipeline (`build_equine_draft` + `render_draft_artifacts` on the real legacy sources and the committed changeset, plus one appended change of `set` ops):
| changes.json | rationale × set ops | Result |
|---|---|---|
| 545 KB | 20 KB × 3,000 | Guard refusal, peak 373 MB |
| 506 KB | 100 KB × 2,000 | Guard refusal, peak 1.19 GB |
| 1.04 MB | 400 KB × 4,000 | `memory allocation failed`, signal 6, at the 6 GB cap |
| 1.92 MB | 80 KB × 14,000 (the case RUST.md declares held) | `memory allocation failed`, signal 6, at the 3 GB cap |
The 16 MiB read cap allows inputs of hundreds of GB of raw table. Without a ulimit the process grows until the OOM killer acts. The failure is fail-closed: `folders()` finishes before `claim_destination`, so nothing gets written. The impact is availability of the developer machine, and that includes `--check`. The test `an_aligned_table_past_the_bound_is_refused_before_it_is_built` calls `format_markdown` on a table it built by hand, so it cannot see the pipeline's allocation.
## Finding 2: records deep-copy whole items
`ChangeRecord` keeps a `ChangeSummary` clone plus deep copies of `before` and `after`, and for a rename or remove those copies are the whole item (`records.rs:18`). An `add` of a 5,000-child category followed by N renames that swap its id back and forth gives:
| renames | build-only RSS |
|---|---|
| 0 | 56 MB |
| 100 | 1.13 GB |
| 200 | 2.20 GB |
| 400 | 4.34 GB (810 KB changes.json) |
`MAX_DOCUMENT_VALUES` is checked on the document after each op, not on what the records accumulate. The cloning lines are R5.3 code, but they sit in a file of this diff (`ops.rs`, whose `apply_rename` the diff modifies), and the R5.4 renderers consume every record. The TypeScript had the same amplification but V8's heap limit stopped it with a clean error. The Rust process has no such ceiling.
## What held
Markdown guard. I ran a differential of `report_guard::format_report` (Rust) against Prettier 3.8.3 `format(…, parser: 'markdown')` with the repository options, on raw reports shaped like the generator's output: list items with title, rationale and question, CHANGELOG rows with `cell()` applied, bibliography and decision lines. Authored texts followed the same preconditions as the code (no backtick; no control character or wiki/liquid bracket in list items). Results:
- 201,000 random probes over the admitted alphabet plus adversarial tokens (`&amp`, `&#`, URL starts, `**`, `_`, `\`, brackets, wide and zero-width characters). 20,969 were admitted and 0 diverged.
- 960 targeted probes: reference-style links with a space, `[a]: b` mid-line, `1.`, `#`, `+`, `>` at a cell start, bare URLs with `_`, `)`, `&`, `x@y`, `**é**`, `a\|b` outside tables, `<3`, `<-`. 689 were admitted and 0 diverged.
None of these forged a section, list item, code span, link or HTML. A newline in a list-item field is refused, and table cells turn newlines into spaces. The guard's scans stayed linear: 18 lines of 4 MB with 20,000 code spans each ran in 1.19 s.
`--check` and `--out`:
- `compare_file` falls back on trees only for `review/app-*` with the declared rule. `compare_drawing` returns an error with no rule, requires the committed file to rebuild byte for byte from its tree, and requires the rule to change it.
- `inside_by_identity` now fails closed with an explicit error.
- Reads stay canonical-inside-seed, regular-file-only and capped at 16 MiB.
- Written names are `Component::Normal` only, and view ids are stable ids.
- Writes use `create_new`, and the existing destination is identity-checked against the seed.
SVG: `drawing_text` refuses unwritable text, and the overlay and the app layer only emit fixed element and attribute names. Depth is bounded by the core reader's prescan before `first_difference` or `TriggerNames::apply` recurse.
## Probes (scratchpad only, no tracked change, nothing deleted)
- `/tmp/claude-1000/-home-bchanot-Documents-faunosteo/9b0edf59-ccd2-4662-b90c-e473dd78028e/scratchpad/probe/src/main.rs` (guard probe), `/tmp/claude-1000/-home-bchanot-Documents-faunosteo/9b0edf59-ccd2-4662-b90c-e473dd78028e/scratchpad/probe/src/bin/amp.rs` (pipeline memory probe; `BUILD_ONLY=1` stops after the build)
- `/tmp/claude-1000/-home-bchanot-Documents-faunosteo/9b0edf59-ccd2-4662-b90c-e473dd78028e/scratchpad/scripts/fuzz.py`, `/tmp/claude-1000/-home-bchanot-Documents-faunosteo/9b0edf59-ccd2-4662-b90c-e473dd78028e/scratchpad/scripts/targeted.py`, `/tmp/claude-1000/-home-bchanot-Documents-faunosteo/9b0edf59-ccd2-4662-b90c-e473dd78028e/scratchpad/scripts/prettier.mjs`
- `/tmp/claude-1000/-home-bchanot-Documents-faunosteo/9b0edf59-ccd2-4662-b90c-e473dd78028e/scratchpad/scripts/mkchg.py` (finding 1 inputs), `/tmp/claude-1000/-home-bchanot-Documents-faunosteo/9b0edf59-ccd2-4662-b90c-e473dd78028e/scratchpad/scripts/mkren.py` (finding 2 inputs)
Memory probes ran under `ulimit -v` (3 to 6 GB). Cargo built into a scratchpad `CARGO_TARGET_DIR`. `git status` stayed clean at `777c800`.
@@ -55,6 +55,11 @@ Fresh verifier and security on the fixed commit: below.
- 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
## Round 7 on 777c800: security BLOCK(2) (verifier pending at fix time); `pnpm check:all` green 14:43:14
- Security (`.claude/audits/2026-10-23-rust-r5-review-security-7.md`): HIGH output amplification: the CHANGELOG raw table (rows × authored rationale) built whole before the round-6 bound (1.9 MB changes.json aborts at 3 GB), RUST.md "refused before any row is built" false; HIGH record amplification: rename/remove records deep-copy the whole item and clone the change texts, no total bound (400 renames of a 5,000-child item → 4.3 GB). LOW test allows, bin allow (context), error echo size, subfolder race (declared out of model).
- Fixed: records bounded in total (`apply::MAX_RECORDED_WEIGHT` 16 Mi units, `records::record_weight`, checked after every op; equine 1,084,422), test `the_records_cannot_outgrow_their_bound`; CHANGELOG tables projected from their cells before the raw table is joined (`markdown::assert_table_bound`), pipeline-level test `a_changelog_whose_alignment_would_explode_is_refused_before_it_is_built`; RUST.md rows rewritten exactly, residual risk (one op's copies before the check).
## Round 8
Fresh verifier and security on the fixed commit: below.