# Narrative sample integration notes

The eight supplied files are retained unchanged. No reuse license was supplied. The supplied test source has not been executed here.

## Browser adaptation

The website's narrative-chunker.ts is derived from src/chunker.ts. It corrects these source-inspected edge cases:

* The oversized-paragraph fallback counted Unicode words but sliced whitespace tokens. One hyphenated token can contain multiple counted words and exceed the claimed hard maximum. The adaptation slices at Unicode word-match boundaries using the same definition as wordCount.
* Sentence boundaries are sliced from the original paragraph so punctuation-only prefixes are not discarded by the original sentence extraction regex.
* Non-finite lexical thresholds and duplicate book/scan-page references are rejected.

The original code remains in this folder for comparison. The browser adaptation and original fictional sample are included under browser/src in the bundle and as public text excerpts.

## Interpretation

This is heuristic scene segmentation, not trained semantic parsing. Same input and options yield deterministic identifiers; those IDs are not persistent across edits. Page references are scan/printed-page provenance, not exact character offsets. Normalization can change legitimate hyphenation; inspect originals. Word counts, scene signals and word-overlap values do not establish semantic accuracy. Large-corpus throughput and memory behavior were not measured. Chunking remains an in-memory operation over the supplied input, not a streaming whole-corpus implementation.

The browser example offers fixed small budgets over the fictional passage and loads only when expanded. It is independent from the workshop OCR review packet and does not silently mix their sources.

No tests, builds, type checks, parser checks, browser checks or sample executions were run during this integration.