perf(common): index artifact bytecode matching

[mattsse]

ContractsByArtifact now keeps exact deployed-bytecode candidates in ordered code-length buckets and lazily prepares each candidate's ignored ranges and unlinked-bytecode representation. The buckets are populated in the existing BTreeMap order, the first fully linked artifact retains a direct fast path, and query metadata is parsed once, preserving the existing earliest-exact and last-partial selection behavior across metadata, immutables, link references, placeholders, and library call protection. This complements #16069's linker path index; it does not change linker resolution.

Candidates store only the artifact id, the match bytecode, and the ignored ranges; a hit resolves back through the artifact map instead of retaining a clone of the contract data, so building the index does not duplicate ABIs or bytecode. That lookup costs roughly 0.3 µs per exact or partial hit and shows up in the small-set middle/last rows below, while misses — which don't touch the map at all — and larger artifact sets stay well ahead of the linear scan.

Results

Criterion 0.7 steady-state medians against master (7c271517a3), same machine and run, using identical 128-length-bucket artifact sets:

Artifacts / query	Candidates checked	master	this PR	delta
100 / first	1 → 1	24.034 ns	17.087 ns	-28.9%
100 / middle	51 → 1	89.220 ns	308.76 ns	+246.1%
100 / last	100 → 1	148.27 ns	481.45 ns	+224.7%
100 / miss	100 → 0	123.53 ns	12.380 ns	-90.0%
1,000 / first	1 → 1	40.046 ns	17.486 ns	-56.3%
1,000 / middle	501 → 4	1.3282 µs	690.39 ns	-48.0%
1,000 / last	1,000 → 8	1.9005 µs	809.66 ns	-57.4%
1,000 / miss	1,000 → 0	1.3922 µs	14.684 ns	-98.9%
5,000 / first	1 → 1	32.689 ns	17.730 ns	-45.8%
5,000 / middle	2,501 → 20	3.9932 µs	749.63 ns	-81.2%
5,000 / last	5,000 → 40	7.9030 µs	1.1539 µs	-85.4%
5,000 / miss	5,000 → 0	7.0576 µs	14.700 ns	-99.8%

The + rows are exact hits in small artifact sets where the final id lookup dominates; the callers (cheatcodes, verify) pay it once per matched query and mostly issue misses against foreign code, which is where the old scan was spending its time.

This PR was written with Codex, including the implementation, tests, benchmarks, and pull request text.

[github-actions]

✅ Changelog found

The deterministic check will validate the changed entry.

[figtracer]

i think we don't need these huge changes for average 100ns save