…das/cuDF (incl. typed edges)

A seeded hop expressed as a gfql()/Cypher chain (not a direct g.hop) always
scanned on pandas/cuDF even with a resident index: _chain_impl attaches a
synthetic per-edge id column (copy(deep=False)+assign) -> a fresh edge-frame
object -> the index's `source_ref is df` identity guard missed.

- registry.rebind_edges: re-point edge adjacency indexes at the augmented frame
  (safe: shallow copy preserves src/dst by value; fingerprint unchanged by an
  added column). Called in chain.py right after the edge-index augmentation.
- Typed edges: relax _hop_is_index_coverable for simple scalar-equality edge_match
  on the wavefront path; index_seeded_hop applies the edge predicate to CSR-matched
  rows each hop, parity-exact with the scan's filter_edges_by_dict.
- Predicate/membership edge_match, edge_query, and the direct-hop path stay on scan.

Measured (dgx, 500k/4M/4-types, warm median, index==scan): pandas typed 2.1x /
untyped 2.7x; cuDF typed 1.9x / untyped 2.0x. polars/polars-gpu untyped already
engage via the native lazy executor; typed polars is a tracked follow-up.

27 new 4-engine differential + engagement tests; existing 109-test index suite green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y6dQEcjdazEnzuvuwf73ZL

… scan path (review blockers)

Review of #1734 found the guard's 'simple scalar equality' narrower than
filter_by_dict's runtime behavior — three defects in one seam, all fixed:
- F1 (BLOCKER): frozenset (+ pd.Index/pd.Series) leaked past the (list, tuple,
  set, dict) check but are membership (isin) on the scan path; a bare == is
  silently all-False -> wrong answer on public g.hop(). The guard now delegates
  to filter_by_dict._is_membership_filter_value (single source of truth);
  EdgeMatchValue widened to match.
- F2 (BLOCKER): null-carrying match columns (pandas nullable dtypes, polars
  nulls — common after NaN->null coercion) produced NA cells whose to_numpy()
  yields an object-dtype array, exploding later at rows[edge_keep[rows]]
  (IndexError). Masks now fill_null(False)/fillna(False) before materializing —
  parity-exact (null == val drops on the scan path too).
- F3 (IMPORTANT): numeric-col-vs-str / string-col-vs-numeric edge_match returned
  a silent empty subgraph where the scan raises (GFQLSchemaError E302 on
  pandas/cuDF; polars its own ComputeError). The mask builder mirrors
  filter_by_dict's exact two checks and declines -> falls back to the scan ->
  users get the SAME error as their engine's scan.
+ 3 engine-parametrized regression tests (frozenset parity, null-column
  no-crash+parity, dtype-mismatch error parity).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y6dQEcjdazEnzuvuwf73ZL