P0: Implement deterministic document-to-chunk capability

[sfloess]

Implement the chunking contract now that FlossWare/scraping#1 is merged.

Acceptance criteria:

• Sentence-aware, approximate token-bounded chunking.
• Configurable overlap and deterministic long-segment splitting.
• Stable chunk object with source identity/provenance, offsets and metadata.
• Preserve upstream metadata.
• Empty input, long segments, overlap boundaries, Unicode/non-ASCII and repeated-run determinism tests.
• Real fixture matching the scraping document contract.
• Core remains stdlib-only and does not import scraping, embedding, storage, model, Loom or agent packages.
• Package/CI validation and X.Y versioning.

Dependency boundary: scraping -> document contract <- chunking.

Program tracker: FlossWare/FlossWare#4

[sfloess]

Implementation order: make this the next P0. Keep the core contract standalone and stdlib-only. Deliver the production implementation + deterministic tests + scraping-compatible fixture + X.Y packaging/CI. Do not introduce Loom/agent/model/storage dependencies. Once the PR is up, it should be reviewed against every acceptance criterion before merge.