# Security audit: R5.6 fingerprint switch, activated (9e87016) Mode audit, fresh and blind. Scope: `git diff 59521dc 9e87016` (31 files, all present at HEAD), branch `feature/rust-r5-pipeline-core`, working tree clean before and after. Contract: `.claude/tasks/contracts/2026-10-25-rust-r5-6-rehearsal.md` (S1-S8 + Clarification 2026-10-25) and the owner's activation brief. No file modified except this report; no network beyond semgrep's rule download; no service key, no Supabase, no transfer tool, no `pnpm format`. ``` SECURITY — VERDICT: PASS TOOL: semgrep 1.168.0 — p/security-audit, p/secrets, p/owasp-top-ten SCOPE: 31 files BLOCKING: (none) REPORTED (non-blocking): - [LOW] rebase-exact-equality-no-sheet-id (pre-existing, R2) — crates/core/src/domain/report/rebase.rs:43 - [LOW] registry-and-declaration-check-public — crates/core/src/domain/sheet/fingerprint.rs:44, :103 - [LOW] snapshot-assets-outside-fingerprint (pre-existing trust model) — crates/core/src/domain/archive/snapshot_rules.rs:79 - [LOW] caller-built-export-input-trusted (pre-existing, documented) — crates/core/src/pdf/write/input.rs:205 - [LOW] new-lint-allow-in-test-file (not a security suppression) — crates/core/tests/r5_6_switch.rs:8 PROOF: semgrep 129 rules on 31 files → 0 findings; checklist 8 checks → 0 findings in scope (+1 guardrails WARN outside the diff); alias-path review 14 checks → 0 blocking, 5 LOW ``` ## Tooling - semgrep 1.168.0, pinned packs, `--metrics=off`: 129 rules run on the 31 changed files, 0 findings, 0 errors. - Guardrails scanner (`~/.claude/skills/guardrails/scripts/scan.py --domain api,supabase`): 0 FAIL, 1 WARN, 1 MANUAL. The WARN (`SB-POLICY-NOT-OWNER`, `supabase/migrations/20261008110000_anatomical_sheets.sql:1177`, `species_select_active`) is in a file outside the diff: out of scope, not counted. MANUAL not reported. - Hermetic tests run in the main tree (no file written outside `target/`): - `cargo test -p faunosteo-core --features cli --locked`: `r5_6_switch` 14/14, `sheet_fingerprint` 11/11 (incl. `nesting_is_bounded_with_a_controlled_error`), `pdf_import` 4/4, `pdf_compare` 16/16; lib unit `the_registered_pair_holds_one_way_for_its_sheet_only` 1/1. - vitest `packages/core-wasm/tests/r5-6.n-1.test.ts` 2/2 and `apps/app/tests/unit/r5-6.switch.test.ts` 2/2 (BUILD_STAMP accepted, module current). - `bash scripts/import-equine.sh --check` (compare-only, read before running): legacy-import 6 files 0 differences, draft 12 files 0 differences. ## Checklist (STEP 3), scoped to the diff | Check | Result | |---|---| | Hardcoded secret / token / key / auth URL | none (added lines grepped; the hex values are public fingerprints) | | SQL by concatenation | none (no SQL in the diff) | | Unsanitised render (innerHTML, eval, raw) | none | | Sensitive endpoint without authz | N/A: no endpoint, RPC or policy in the diff | | Stack trace / internal path to the user | none new. `facade.rs:56` maps `LoadError::Fingerprint` to `invalid_sheet` with the refusal's Debug text (a recomputed hex or a serde reason about the page's own bundle), no path, no secret | | Secret / PII in a log | none (no log call added) | | Tracked `.env` / credential file | only `.env.example` tracked | | Guardrails scanner | run, nothing in scope (see above) | Anti-gaming (STEP 4): no new `nosemgrep`, `nosec`, `eslint-disable`, `@ts-ignore` in the diff. One new `#![allow(clippy::disallowed_methods, clippy::unwrap_used, clippy::expect_used)]` heads the new test file `crates/core/tests/r5_6_switch.rs:8`. It is the shared header of 20+ integration test files (the core's no-I/O and no-panic lints, which tests that read fixtures must relax), not a security-rule suppression: reported LOW, not blocking. ## Alias paths attacked (the brief) 1. **`equivalence` crate-private.** `fingerprint.rs:72` is `pub(crate)`. Its one non-test caller is `LoadedSheet::recognises_historical_stamp` (`loaded.rs:169`). The wasm crate references no fingerprint item (`crates/wasm/src` grep empty). 2. **One decision point.** Every alias decision goes through `recognises_historical_stamp` (`loaded.rs:163-171`): `build_sheet_snapshot` (`snapshot_rules.rs:62-64`, behind `archive.snapshot` `ops.rs:361-363`), `rebase_report_on_sheet` (`rebase.rs:43-46`, reached by `session.rs:89` draft restore and `session.rs:127` import), `sheet_line` (`compare/lines.rs:350-358`). The only other reader of `TRUSTED_ALIASES` in `src` is the generator (`pipeline.rs:185`), which writes the descriptive list into `sheet-ref.json`; that list is never read back for a decision (`declared.aliases` is only passed to `check_declared`, `loaded.rs:199-204`). 3. **Forging the private verified state.** `verified_fingerprint` (`loaded.rs:75`) is a private field; `from_bundle` is the only constructor (`loaded.rs:121-151`), `LoadedSheet` derives neither `Deserialize` nor `Default`, struct-update syntax is impossible outside the module, and `unsafe_code = "forbid"` applies to core and wasm (workspace lints, `Cargo.toml:55`, both crates `[lints] workspace = true`). `Clone` copies the state with the document, which the recomputation then re-proves. The wasm `SheetHandle` keeps its `LoadedSheet` private (`crates/wasm/src/lib.rs:27`), so a page cannot mutate it. 4. **Recomputation at the decision.** The predicate requires a verified value, `verified == reference.document_sha256`, the registry pair `(sheet_id, alias → verified)` in one direction, and `document_fingerprint(&self.document) == verified` now. A mutated pub `document` or `reference`, a cloned sheet re-pointed to another sheet id, a v0 sheet, or a stamp the registry does not bind all return false (tests `a_mutated_document_or_reference_never_keeps_the_alias`, `the_historical_value_on_another_sheet_id_is_not_recognised`). A bundle with another sheet id and the genuine document verifies but never matches the registry (entry pinned to the equine id); a bundle with the equine id and other content verifies to another value the registry does not bind. Granting the alias therefore needs the genuine content (SHA-256 preimage otherwise). 5. **Injectivity of the fingerprinted form.** The value hashes the typed reading serialised back. Checked for collapses that would let two behaviours share a value: no `#[serde(skip)]` or `flatten` in `sheet/schema.rs`; `skip_serializing_if = "Option::is_none"` (lines 169, 176) round-trips; the untagged enums (`SheetView`, `Geometry`) carry distinguishing fields. The app consumes the same typed document, so text-only differences the reading erases are never behaviour. 6. **v0 / v1 and direction.** No block = v0, unverified, no privilege (`check_reference` returns `Ok(false)`, `loaded.rs:196-198`); an unverified v0 sheet carrying the canonical value still rebases an old report and its direct export is refused (`SheetMismatch`, test `the_direct_export_on_an_unverified_v0_sheet_with_the_canonical_value_is_refused`). The registry is consulted only as alias → canonical, so a v0 (rollback) sheet never recognises a `v1` stamp. `check_declared` refuses a declared value that is a registered alias or not 64 lowercase hex (`fingerprint.rs:112-118`). 7. **`null` and malformed blocks.** `present_block` (`loaded.rs:46-50`) turns an explicit `null` into a shape error; `DeclaredFingerprint` is `deny_unknown_fields` with required `convention` and `aliases` (`loaded.rs:52-57`); a string block and an unknown key are refused (tests). A misspelt key on the non-strict `BundledRef` only downgrades to v0: fail-closed, no privilege. 8. **Aliases never from a report, archive or bundle.** Archives cannot carry one: `SheetSnapshotRef` and the snapshot are `deny_unknown_fields` (`archive/snapshot.rs:42, 59`); the archive reader keeps strict equality between snapshot and report stamps (`archive/read.rs:116-119`, unchanged). 9. **Export input strict.** `assert_exportable` (`pdf/write/input.rs:204-210`) is unchanged by the diff: strict sheet id and stamp equality, no registry. The alias only reaches an archive through the core-built snapshot of a verified sheet. 10. **`report:compare`.** The alias line prints both raw values truncated to 12 chars through `esc` (`compare/lines.rs:366-376`); `same_sheet` still requires the sheet id; the bin reads the seed's block through the same `present_block` and `from_bundle` verification (`bin/report_compare.rs:40-44, 66`). 11. **`canonical_json` depth.** Bounded at `MAX_CANONICAL_DEPTH = 128` with `CanonicalError::TooDeep` (`fingerprint.rs:132-154`), above serde_json's parse limit (127 nested containers accepted), so every parsed bundle fits and a value built in code beyond it errors instead of overflowing the stack. Tested at 128 / 129, arrays and objects. 12. **Recomputation cost.** In the predicate the recomputation is the last conjunct, after the registry match: a forged or unknown stamp costs a string compare, only the registered historical stamp triggers one recomputation (one per export snapshot, rebase or compare line). The load recomputes once per `SheetHandle` when a block is present (page runtime, `runtime.ts:322`, and each import worker, `handler.ts:48`), O(document), 302,794 B for the equine draft. Not attacker-amplifiable: the bundle is build-time, a report triggers at most one recomputation per call. The 14 switch tests (loads, imports, exports) run in 0.81 s in a debug build. 13. **Seed artifacts.** Only `draft/sheet-ref.json` and `draft/report-migration-map.json` change. `document.json` (sha256 27c7383c…), `assets/horse.svg` (0ea5202f…) and `equine_draft.sql` (c5a24312…) hash identically at 59521dc and HEAD. The committed block is consistent: value `12d78872…`, convention `v1`, alias `27c7383c…`, which is the registry entry and the old file-bytes hash. The migration map changes only its `documentSha256`. 14. **Oracles, goldens, N-1.** 0 files changed under `crates/core/tests/oracles`, `crates/core/tests/pdf`, `apps/app/tests/fixtures`, `packages/core-wasm/tests/n-1`. The new `crates/core/tests/fixtures/r5-6/pre-switch/{sheet-ref,report-migration-map}.json` are byte-identical to the 59521dc seed files (sha256 2c0716f1…, 4139c51d…), so the redirected historical tests read the real pre-switch inputs and cannot mask an anatomical change. Exact-equality semantics are unchanged and documented as no proof of content (`fingerprint.rs:13-18`). ## Reported findings (LOW, non-blocking) **L1. Rebase exact-equality branch has no sheet-id check** — `crates/core/src/domain/report/rebase.rs:43`. Pre-existing since R2, unchanged in meaning: a report of another sheet id stamped with the current sheet's value skips the rebase. The new alias branch does check the sheet id. Hint: add `report.sheet.sheet_id == sheet.reference.sheet_id` to the equality branch in a later step (behaviour change, owner decision). **L2. Registry and declaration check stay public** — `crates/core/src/domain/sheet/fingerprint.rs:44` (`TRUSTED_ALIASES`), `:103` (`check_declared`). Neither grants anything by itself and no consumer misuses them today (integration tests and the generator need them), but a future crate consumer could rebuild an alias decision from the registry without the recomputation. Hint: `#[doc(hidden)]` plus a doc line, or move the tests in-crate and make both `pub(crate)`. **L3. Archived drawings are outside the document fingerprint** — `crates/core/src/domain/archive/snapshot_rules.rs:79` (`snapshot_assets(sheet)`). A hand-made bundle with the genuine document but other SVG texts would archive them under the historical (or the canonical) stamp. Pre-existing trust model (the bundle is trusted for what it declares, `fingerprint.rs:15-18`); the alias adds no privilege the canonical stamp did not already have. Hint: bind asset hashes into the document or check `asset_hashes` against the texts at load, if the threat model ever includes non-build bundles. **L4. Caller-built export input trusted for its self-consistent report and snapshot** — `crates/core/src/pdf/write/input.rs:205`. A page script can hand the export any matching pair; the registry is never consulted there and every reader decides again on import. Documented residual, unchanged by the diff. **L5. New lint allow in a test file** — `crates/core/tests/r5_6_switch.rs:8`. Shared test-file convention for the core's I/O and panic lints, not a security suppression. SECURITY — VERDICT: PASS