fix(relay): stop overcounting NIP-45 COUNT for non-pushable filter constraints

[1nderboi]

Summary

Fixes #3373. filter_fully_pushable() was letting several filter shapes through to the fast SQL COUNT path that filter_to_query_params() actually drops instead of pushing as predicates — causing COUNT to overcount:

• A present-but-empty constraint (authors: [], ids: [], or any generic tag with zero values) matches nothing under filters_match, but the query builder just omits it rather than emitting a match-nothing predicate.
• A single #h value that isn't a channel UUID (e.g. a bare NIP-29 group id) was treated as pushable, but only UUID-valued #h is pushed as channel_id — a non-UUID value needs post-filtering since events can legitimately carry it as a literal tag.
• Any other generic tag (#t, #a, etc.) with a non-empty value list was allowed through when it should always require post-filtering.

All three now correctly fall back to the exact post-filtered COUNT path.

Changes

• crates/buzz-relay/src/handlers/req.rs: reject empty authors/ids/generic-tag constraints, reject non-UUID single #h, reject any non-empty non-h/p/d/e generic tag.
• 6 new unit tests covering: UUID vs non-UUID single #h, multi-#h, empty authors/ids/generic-tag constraints, non-empty #t, and the plain kind+author baseline that should stay pushable.

Test plan

• cargo fmt -p buzz-relay --check — clean
• Manual trace of filter_fully_pushable against filters_match / filter_to_query_params for each new case
• cargo test -p buzz-relay — could not run to completion locally. This shared dev machine's disk repeatedly hit ENOSPC mid-build (down to 0 bytes free at one point) even after cargo clean-ing target dirs; the buzz-relay dependency tree (aws-lc-sys, sqlx-postgres, opentelemetry_sdk, mesh-llm-native-runtime) alone exceeds available local headroom. Requesting CI run the full suite including the new tests in req.rs's mod tests.

🤖 Generated with Claude Code

[1nderboi]

Closing in favor of #3374, which fixes the identical bug (filter_fully_pushable letting a non-UUID single #h and any present-but-empty constraint through to the fast COUNT path) the same way, was opened first, and additionally has an e2e wire-level regression test against a real relay (test_count_unpushable_filters_do_not_count_all_channels) on top of unit tests — stronger coverage than this PR's unit-tests-only approach.

For the record: locally confirmed this branch's fix works too — cargo test -p buzz-relay --lib req:: → 51 passed, 0 failed — but no need for two competing fixes on the same issue. Thanks @beardthelion for #3374.