feat(ingest): LLM-driven TOC tree builder (PageIndex-style)

Ports the three-phase PageIndex pipeline into Go:

  1. detect    — single-page TOC detector over the first N pages.
  2. extract   — if a TOC page was found, parse it into nested
                 nodes; otherwise fall through to the no-TOC
                 path that generates a TOC straight from body
                 text tagged with <physical_index_X> markers.
  3. verify    — concurrently re-check every node's claimed
                 start page; mismatches clear the page back to
                 zero (the "unknown / open" sentinel) rather
                 than making one up.

End pages are derived from sibling ordering after verification.
Node IDs are stamped deterministically from the dotted
structure so callers can diff trees across re-ingestions.

The retry-on-parse-failure helper mirrors
pkg/retrieval/single_pass.go::runSelectionWithRetry; it is
duplicated rather than imported so the builder doesn't drag the
retrieval package into its dependency graph.

LLM parse blips degrade to "no usable nodes" with a logged
warning so a single bad response never fails ingest — the
document remains fully retrievable via the existing sections
tree.

Tests cover the happy path, the no-TOC path, verification
repair, JSON retry, end-page derivation, hierarchy assembly,
the <physical_index_X> tag parser, and the empty-input
short-circuit.

Author: Halleluyah Oladipo