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

[Felipe705x]

Mergeable port of the pinned-baseline benchmark (#93) onto current main: identical instrument, cases, and protocol; imports under the current lib name. #93 stays open-unmerged as the fixed v0.2.0 reference; this one lives with the code going forward — after any future checker change, running both gives a before/after scaling comparison on the identical instrument.

• 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); the chain is parameterized in k (1–16) by inserting intermediate nodes only.
• Schema family: homogeneous, parameterized by total entry count S; 1:2 node:edge ratio matching the measured LDBC inferred schema (36 entries, identical in SF0.1 and SF0.3 — the inferred schema is scale-invariant, hence the synthetic family). Suggested ladder: 16–1024 (powers of two) + 36 as the LDBC anchor.
• Protocol: parse/elaborate outside the timed loop; median of ≤4000 iterations under a 200 ms/case budget; verdicts asserted every run; ≥3 interleaved sweeps aggregated by medians.

Smoke-tested on this branch (all verdicts correct). cargo clippy clean on the new bin.

🤖 Generated with Claude Code