Correct eleven gold annotations pointing at dead, shim, or deprecated files

[Hartistic]

Follow-up to #257 — the corrections offered there, folded into one PR.

The benchmark scores a tool by whether it finds one specific "right answer" file for each question. For these eleven questions, the file the answer key names isn't where the behavior in the question actually lives: some of the named files hold copies of code nothing in the project calls anymore, some are small forwarding files that only point at the real implementation, and one is an entry point the project itself now warns against using. A tool that returns the genuinely useful file gets marked wrong on these; one that returns the leftover gets full credit. Each replacement below was checked by hand at the revisions pinned in repos.json (call-site checks and file reads — the per-line evidence is in the issue thread).

repo	query	was	now
model2vec	how vocabulary is pruned during distillation	distill/utils.py (device selection only)	tokenizer/tokenizer.py, where prune_added_tokens() lives (secondary distill/distillation.py unchanged)
model2vec	tokenizer construction and vocabulary building	secondary distill/utils.py	secondary dropped
pydantic	custom field and model validators	secondary class_validators.py (5-line migration shim)	secondary deprecated/class_validators.py, the actual V1-style validators
aiohttp	WebSocket client implementation	secondary _websocket/reader.py (31-line C/Python import switch)	secondary _websocket/reader_py.py, the implementation it re-exports
vitest	test reporter interface for listening to test events and results	public/reporters.ts (re-export barrel, deprecated at runtime since Vitest 4.1)	node/types/reporter.ts (the Reporter interface itself), secondary node/reporters/index.ts

The six zod queries (union/discriminatedUnion, optional/nullable, transform/pipe, record/map, ZodDefault/ZodCatch, enum/nativeEnum/literal) all share one pattern: v4/core/api.ts exports factory helpers that classic/ and mini/ reimplement rather than call — 49 of its 114 exported functions have no call site anywhere in the repo. Each of the six now points at v4/core/schemas.ts (the runtime classes that do the work) as primary, with v4/classic/schemas.ts (the builders users actually reach through z.*) as secondary.

Reran semble itself on the five affected repos before and after, same machine, scored the same way as the head-to-head in the issue:

repo	n	before	after	Δ
aiohttp	21	0.773	0.762	−0.011
model2vec	20	0.716	0.738	+0.022
pydantic	20	0.690	0.690	0.000
vitest	20	0.654	0.641	−0.013
zod	20	0.558	0.639	+0.081
overall	101	0.679	0.695	+0.016

Net positive for semble itself, mostly from zod — pointing the golds at files that actually contain the queried behavior makes them more findable for everyone. I didn't rerun the other models in the published table (as you said, they're the expensive part), so their columns will drift a little on these repos whenever they are.

🤖 Generated with Claude Code

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge; no concrete incorrect target, malformed annotation, or benchmark-contract violation was identified.

The changed files retain the expected annotation structure, and the new targets align with the substantive implementations or definitions described by each query without an established scoring failure.

_{Reviews (1): Last reviewed commit: "fix(benchmarks): correct eleven gold ann..." | Re-trigger Greptile}