Synthetic malformed header fixtures¶

The multimodal preflight suite already refuses malformed document and medical-image headers before any decode. This contribution adds reusable, deterministic corruption fixtures and focused regression tests for those boundaries; it does not add PDF or DICOM parsing support, relax any refusal, or change runtime behavior.

from openmed.multimodal.pdf_geometry import read_pdf_geometry
from tests.fixtures.multimodal.malformed import case_by_name

case = case_by_name("pdf-header-cyclic")
report = read_pdf_geometry(case.payload)
assert report.reason_codes == ("pdf_page_tree_invalid",)

tests/fixtures/multimodal/malformed.py exposes one table, MALFORMED_HEADER_CASES, with five corruption classes for each modality: truncated, oversized, cyclic, inconsistent, and unsupported. Each entry records the payload, the boundary that refuses it, and the exact reason that boundary reports, so tests and documentation cannot drift from implemented behavior.

Pre-decode boundaries¶

PDF fixtures are consumed by openmed.multimodal.pdf_geometry.read_pdf_geometry, the bounded, dependency-free header and page-tree reader. Corrupt input does not raise: the reader returns a report whose reason_codes are stable members of PDF_REASON_CODES.

pdf-header-oversized is a twelve-object file probed with an explicit max_objects=4 budget, so the fixture exercises the reader's bounded object allowance without carrying a large payload. pdf-header-inconsistent is the one case that is not rejected: a declared /Count 3 over a single page-tree leaf is reported as page_count_mismatch on a review verdict, which is the reader's existing precedence rule.

DICOM boundary limitation¶

This repository has no byte-level DICOM structure preflight. The only raw byte boundary is the media-type check in openmed.multimodal.media_type.detect_media_type, which inspects at most MAX_MEDIA_TYPE_PREFIX_BYTES (132) leading bytes and matches the DICM magic at offset 128, followed by openmed.multimodal.preflight.preflight_asset, which reports unknown or mismatch under its media_type check. Every DICOM decode path takes a path and calls pydicom.dcmread(..., force=True), so a truncated or badly sized data set may never raise there.

The DICOM fixtures therefore document the bounded magic boundary and the preflight media-type verdict only. They never reach a decoder and make no claim about group length, transfer syntax, or pixel data.

Each DICOM fixture also records the manifest declaration it is compared against (a frames/width/height DICOM manifest, or an image/bmp manifest for the inconsistent case), so the preflight report contains exactly one finding: the media-type verdict that classifies the corruption.

Determinism and privacy¶

Every fixture is assembled from constants only — no randomness, no timestamps, no third-party dependency. Payloads are at most 4096 bytes and, apart from the PDF structural dictionaries, contain no text, no patient identifiers, no clinical pixels, and no encoded image data. PAYLOAD_SHA256 pins each payload's digest, and the tests fail if a fixture's bytes, size, or recorded reason code drift.

Verification¶

uv run --frozen --extra dev pytest tests/unit/multimodal/test_malformed_headers.py -q

The tests run entirely offline and assert, per fixture, the boundary's status, the exact reason code, the media type reported by the bounded sniffer, the presence of a valid DICM magic only where one is expected, and the absence of PHI markers. Existing PDF geometry, media-type, and preflight tests are retained unmodified. No network calls, decoders, model weights, or credentials are needed.