Reject a non-null Prev entry that cannot be read as an offset instead of silently treating it as the end of the guarded trailer chain. Keep absent/null end markers and readable numeric offsets unchanged.
Reproduced CodeRabbit discussion_r4366245854 with an independent probe and a failing native regression before the fix. The synthetic regression covers seven unreadable value types plus absent/null and integer/real history controls.
Validation: all 9 cargo xtask ci gates passed (3063 tests passed, 6 skipped; 16 WASM-safe crates). One unrelated GPU test was labelled leaky by nextest in that run and passed cleanly in an isolated rerun. Independent checks: 17 original supported PDF variants still pass qpdf, logical preservation and Poppler pixels; 6 original refusal cases remain refused. All 11 additional Prev controls pass (7 refused, 4 accepted).
Integrate current main (d8d78851) without rewriting the contributor's commits, retaining craft-libs 9f75d3dc.
Accept literal null trailer entries while rejecting actual or ambiguous encryption/hybrid markers. Structurally inspect bounded xref/trailer history instead of adopting trailer-looking comments. Preflight all selected source variants before editing/undo; preserve reset-to-defaults. Suppress stale unrewritable canvas overrides rather than revealing source content. Keep conservative all-content crop fallback without overcropping layerless/null stubs.
Validation: cargo xtask ci, all 9 gates passed; 3062 tests passed, 6 skipped; all 16 WASM-safe crates checked. 17 supported variants passed independent qpdf, logical preservation and Poppler pixel checks; 6 unsupported/repair-only cases refused. Direct-harness Claude Opus 5 reviewed the exact main-relative delta with no blocking findings.
No native-format changes or unrelated layer features. Synthetic regression fixtures are generated in test code.
