bench: typecheck_scaling — schema-size × query-size scaling of the type checker

[Felipe705x]

Do not merge — pinned-baseline benchmark. This branch deliberately sits on the v0.2.0 release (6cba43e) so the numbers track a fixed released baseline of the checker; the bin uses the crate's lib name at that point (gqlrust), so merging into main (lib renamed to frogql) would break the build. Run the bench from this branch; cite it as bench/typecheck-scaling @ HEAD.

Runtime-independent microbenchmark of Typechecker::check_query: cost as a function of schema size (a homogeneous synthetic schema family — the controlled variable, so a fixed real schema can't serve) and query size (chain length k). No runtime, storage, or data involved.

• Cases: from internal_bench.rs — v_chain_knows, e_bad_edge_deep, e_label_only — byte-identical after renaming labels into the synthetic universe (Person→L0, knows→e0, Wagumi→NoSuchLabel; projected property names exist verbatim). The chain is parameterized in k by inserting intermediate nodes only; k=1 is the original query.
• Schema family: max(S/3,2) node types (two properties each) + undirected edge types, 1:2 ratio matching the measured LDBC inferred schema — 11 node + 25 edge entries = 36, identical for SF0.1 and SF0.3 (verified on both datasets; the inferred schema is scale-invariant, which is precisely why varying schema size requires a synthetic family). Suggested ladder: 16–1024 (powers of two) plus 36 as the LDBC-size anchor.
• Protocol: parse/elaborate outside the timed loop; median of ≤4000 iterations under a 200 ms/case budget; verdicts asserted every run. Aggregate ≥3 interleaved sweeps by medians (session-to-session wall-clock varies on consumer hardware).

🤖 Generated with Claude Code