Phase 6 follow-up: GROBID processFulltext fallback for TOC-less book PDFs

[deOliveira-R]

`extract.book_structure.detect` splits books using the PDF's embedded TOC (`pymupdf.get_toc()`) or the EPUB spine. When a PDF has neither — scanned or badly-produced files — the book stays as a single item and the "chapters as first-class items" feature (plan §6 Phase 6a) silently skips it.

Plan hint at the time:

PDF second (pymupdf bookmark tree; fall back to GROBID's `processFullText` when bookmarks are missing)

Proposal

1. When `detect()` returns `None` for a PDF, optionally call GROBID's `/api/processFulltextDocument` (already wrapped via `extract.grobid` — only the header endpoint is used today).
2. Parse the TEI response for `<div type="chapter">` or top-level `` structures.
3. Convert to `ChapterSpec` entries.
4. Opt-in via a config flag (`GRIMOIRE_GROBID_FULLTEXT_FALLBACK=true`) — the endpoint is 3–10 s per PDF and not always reliable.

Acceptance

• `detect()` honors the opt-in config flag and falls back to GROBID only when TOC + spine detection both returned `None`.
• Oracle: add a TOC-less book PDF fixture (or synthetic) and assert chapters are created when the flag is on.
• Failure modes (GROBID down, malformed TEI, no `<div type="chapter">` structure) degrade gracefully — book stays as a single item, no partial chapters.