…1ms (#1755 lever-3)

Adds _seeded_typed_hop_numpy_pandas: a pandas-only numpy specialization of the
seeded typed 1-hop, gated to scalar node/edge filters + directed (falls back to
the engine-agnostic seeded branch for predicates/undirected/cuDF). Seed-reduces
edges before the object-typed edge_match scan, applies the destination filter to
the small gathered dst nodes, and does a SINGLE full-node scan to gather result
nodes — so a seeded lookup is a few numpy passes, not O(E) pandas machinery.

Measured dgx (26.02-gfql-polars, 50k/200k, native [n({id}),e(edge_match),n({type})],
warm-median): full path 29.10 ms -> fast path 0.553 ms (52.6x), byte-identical
parity. GATE MET (<=1ms) on the NATIVE seeded typed hop. All 57 chain tests pass,
mypy clean.

NOTE: this closes the native surface only. The cypher string
g.gfql("MATCH (m {id})-[:T]->(p) RETURN p") lowers to NAMED nodes (m,p) +
label__ filters + a 4th ASTCall projection, so it does not yet engage this fast
path (63 ms). Closing the cypher surface = the next lever (named-node + label +
lean projection). Prototype on the bench branch (fast-path infra not yet on master).

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

…<=1ms (#1755)

Closes the #1755 seeded-Cypher abstraction tax on the CYPHER STRING surface.
g.gfql("MATCH (m {id})-[:T]->(p) RETURN p") lowered to named nodes + label__
filters + a rows/projection pipeline that scanned all edges/nodes and paid ~30ms
of rows-pivot machinery even for a single-row seeded result.

New _execute_seeded_typed_hop_fast_path (gfql_unified, alongside the existing
single-hop-aggregate / two-hop-count fast paths): detects the seeded typed 1-hop
with a single whole-row node RETURN, then:
 - _seeded_typed_return_dst_pandas (chain.py): numpy dst-only extraction — resolves
   label__X->type, id-first successive subset of the seed (later filters index the
   tiny survivor rows via .iloc so object columns are never fully materialized),
   seed-reduces edges before the typed-edge scan, gathers + filters the dst nodes
   in one node pass;
 - a lean apply_result_projection on just the p rows (exact column-order/flatten
   semantics, no rows-pivot).
Gated pandas-only, forward seeded shape (RETURN alias == dst node, seed on src);
reverse / multi-alias / field / non-scalar returns fall through to the full path.

Parity byte-identical fast-vs-full across RETURN p (labelled/unlabelled dst),
no-match, and the bail shapes. Measured dgx (26.02-gfql-polars, 50k/200k,
warm-median): full 38.2ms -> fast 0.847 ms (45.2x). GATE MET (<=1ms) on the
seeded cypher string. 57 chain tests pass, mypy clean.

Also lands the native lever-3 helper (_seeded_typed_hop_numpy_pandas + the
_try_chain_fast_path seed-first typed-edge path) this builds on: native seeded
typed hop 29.1 -> 0.553 ms.

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

…1755)

Re-verify the seeded-Cypher tax fast path against an INDEPENDENT oracle
(creators hand-computed from raw edge/node frames), not merely fast==full,
since the full path could share a bug. Native returns {seed} u creators
(both bound endpoints); cypher RETURN p returns creators only.

Decline coverage: two-hop, predicate node filter, predicate edge filter,
undirected -> fast path declines and the full path result is byte-identical.

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

…1755)

The cypher seeded fast path gated on isinstance(e1, ASTEdge) but not on the
edge being a genuine single hop. A variable-length edge (-[*1..2]->) is one
ASTEdge spanning multiple hops, so the 1-hop seed reduction silently truncated
it — MATCH (a {id})-[*1..2]->(b) RETURN b returned only the 1-hop neighbors,
breaking pandas-vs-polars parity (TestVarlenAliasHopGate).

Gate on e1.is_simple_single_hop() (same canonical check the native fast path
uses; also rejects hop labels, output slicing, fixed-point). Adds varlen
decline coverage to the fast-path parity suite.

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

Annotate _seeded_typed_hop_numpy_pandas / _seeded_typed_return_dst_pandas
(Plottable/ASTNode/ASTEdge params, Optional[Plottable] and
Optional[Tuple[DataFrameT, DataFrameT]] returns) so their bodies are
type-checked, and guard the src/dst/node bindings against None at the
_try_chain_fast_path call site. mypy --config-file mypy.ini clean on both
changed files (remaining errors are pre-existing polars-lazy, untouched).

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

…#1755)

Review feedback on the seeded typed-hop fast path:
- direction: str -> Direction (the canonical Literal from ast.py) in both helpers.
- Make the helpers ENGINE-GENERIC (pandas + cuDF): _seeded_typed_hop_numpy_pandas
  -> _seeded_typed_hop_pandas_cudf and _seeded_typed_return_dst_pandas ->
  _seeded_typed_return_dst_pandas_cudf, rewritten to use only the shared
  pandas/cuDF DataFrame API (no numpy array drops), so the same body runs on
  both engines. Relax the native (_try_chain_fast_path) and cypher
  (_execute_seeded_typed_hop_fast_path) gates to Engine.PANDAS|CUDF.
- Replace dynamic getattr with static attribute access in gfql_unified
  (projection.table/.columns, proj_cols[0].kind, n2._name, e1.direction,
  base_graph._node/_source/_destination) now that types are statically known.
- Drop the dead _SEEDED_FASTPATH_GUARD marker (never set; vestigial from the
  earlier re-run design) — removes the last getattr.

Pandas parity + independent-oracle tests still byte-identical (19 pass); mypy
clean on both files (only pre-existing polars-lazy errors remain). cuDF parity
verified separately on GPU.

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

…1755)

Answers the coverage question: broaden ENGAGED-shape verification beyond the few
hardcoded cases. Adds a 7-shape × {pandas, cuDF} matrix asserting fast-on ==
fast-off byte-identical (native typed/untyped/reverse/type-src + cypher
labelled/unlabelled/no-match), plus an independent-oracle check per engine. cuDF
cases importorskip locally, run on GPU/CI. Complements the ~1114 implicit
decline-path exercises the cypher conformance suite already drives through the
dispatch (real queries that fall through + still return correct values). Also
makes _canon_nodes engine-aware (cuDF/polars -> pandas via to_pandas).

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

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