Ordered multimodal batch preflight¶

preflight_asset_batch composes the existing preflight_asset contract over ordered validated manifests and caller-owned local sources. It preserves input positions rather than applying the inventory AssetBatch canonical-ID sort. It does not decode media, load a model, choose a provider or delete an asset.

import hashlib
from openmed.multimodal.asset_manifest import AssetManifest
from openmed.multimodal.batch_preflight import (
    BatchPreflightAsset, BatchPreflightLimits, preflight_asset_batch,
)

header = b'\x89PNG\r\n\x1a\n'
manifest = AssetManifest(
    'synthetic_image', 'image/png', hashlib.sha256(header).hexdigest(),
    len(header), width=8, height=8,
)
limits = BatchPreflightLimits(
    bytes=1024, pages=10, pixels=10000, frames=10, duration=10, rejections=0,
)
report = preflight_asset_batch([BatchPreflightAsset(manifest, header)], limits=limits)
assert report.outcome == 'accept'
print(report.to_json())

Aggregate bytes, pages, pixels, frames, duration and rejected positions are checked against explicit inclusive ceilings. Integer totals have the existing manifest/pixel-product bounds; duration uses deterministic math.fsum. Overflow or missing applicable geometry/duration produces a finding and a None total, never an assumed zero. PDF raster geometry is not represented by the existing preflight contract: PDFs retain its abstention and hold the batch. This does not newly claim that arbitrary signature fragments are decodable media.

Each asset has exactly one ordered result: accept, abstain, error or not_evaluated. Accepted and abstaining assets retain their actual shared preflight report without rewriting it. Duplicate IDs and content digests, aggregate failures and source-read errors produce independent ordered findings. Source bytes, stream representations and underlying exception messages are excluded from results; sources stay caller-owned and are never closed.

• continue examines subsequent assets admitted by the aggregate byte ceiling, retaining all findings and metadata. It does not override rejection.
• hold stops source reads at the first rejected position and holds the packet.
• fail_fast stops source reads at the first rejected position and returns the explicit fail-fast outcome.

Aggregate and duplicate metadata checks continue across all submitted manifests in every mode. Exceeding or overflowing the aggregate byte ceiling prevents source reads even in continue mode. Any finding prevents accept; a continue report with a rejected asset therefore has outcome hold. Later stopped assets remain not_evaluated. At most 10,000 assets are consumed, with one additional probe to detect an over-limit iterable.

Reports are deterministic JSON containing checked resource counts, policy ceilings and existing privacy-safe manifest/finding metadata only. Input text, transcripts, DICOM values, credentials, paths and raw source bytes are excluded. Preflight acceptance is resource/metadata admission, not clinical validity, packet completeness or proof that subsequent redaction succeeds.