Processing Pipelines

Importing a script runs a four-stage pipeline that turns raw PDF bytes into a structured show.

PDF bytes
   │
   ▼
[1] Ingest (crates/ocr)      extract_text()  ── native text layer, else OCR
   │  plain text, one line per source line
   ▼
[2] Classify (crates/classify) classify_line()  ── Heading / StageDirection / Dialog / Noise
   │  Vec<ClassifiedLine>
   ▼
[3] Construct (crates/show)   construct_show()  ── build Section/Cue tree
   │  Show (raw)
   ▼
[4] Filter (crates/show)      filter_show()      ── drop noise & empty sections
   │
   ▼
Show

show::pipeline::build_show_from_text runs stages 2–4; import_pdf_show runs stage 1 and then calls it.

Stage responsibilities

Stage	Crate / file	Does
Ingest	ocr/src/pipeline.rs	PDF → text. Detects native text; else renders + OCRs.
Classify	classify/src/rules.rs	Each line → a LineKind (and headings → HeadingKind).
Construct	show/src/pipeline.rs	Assembles classified lines into acts/scenes/cues.
Filter	show/src/filter.rs	Removes noise cues and (optionally) empty sections.

Why the stages are separate

Each stage has a single responsibility and is independently testable:

• Ingest only gets text out of the document. It never interprets structure.
• Classify is a pure function from line to label, so the rules can be refined and unit-tested without touching tree-building.
• Construct follows structural rules and can be tested with synthetic classified input.
• Filter is a cleanup pass that can be relaxed or extended without disturbing construction.

Continue with OCR ingest, Classify, and Construct & filter.