digest-pdf — digest PDFs into a structured (image + text) tree

You orchestrate the digest-pdf tool. The tool does all recognition/cutting with zero LLM tokens (MinerU is the sole transcriber, ADR-0010); your job is to understand the invoking repo, confirm its Layout Spec (where Units land — CONTEXT: Layout Spec, ADR-0008) once, run a fast batch, and verify. Never OCR or transcribe pages yourself by reading images — that defeats the tool's design and is slow/costly.

Inputs (parse from the user's message)

• targets — one or more PDF files and/or directories (the tool recurses dirs for *.pdf).
• --out — base override for the output root. Usually omit it: when the invoking repo has a Layout Spec, the tool lands each PDF per its path template resolved from the repo root. Pass --out only to write outside the repo (or on a no-spec run, where each PDF lands under <out>/<pdf-stem>/).
• --strategy — auto (default) | page | outline | question. A matching Layout Spec rule pins the strategy per PDF (overriding this); pass --strategy only on a no-spec run.
• --layout — path to a Layout Spec; defaults to auto-discovering .digest/layout.yaml by walking up from cwd.