feat: add NIP-91 AND filter support

[hzrd149]

Summary

• implement draft NIP-91 &x tag filters with all-values-required semantics
• normalize compatibility #x filters so AND values do not satisfy the residual OR clause
• use one required tag as the historical/live index seed, then verify the complete packed event
• support NIP-91 across REQ, live delivery, search, COUNT, router, sync, and negentropy
• advertise NIP-91 and configurable max_and_entries/max_tags_per_filter limits through NIP-11

The implementation follows nostr-protocol/nips#2252 at b93bda29d45998866e81c65e0693616294a78672 and takes the overlap-normalization approach from hoytech/strfry#162.

Safety and compatibility

• reject malformed, empty, oversized, and invalid hex AND filters
• force full-event verification rather than unsafe index-only delivery
• omit NIP-45 HLL sketches for ambiguous AND filter shapes
• preserve restricted COUNT privacy rules
• post-filter results received from relays that only understand the compatibility # clause
• keep persisted negentropy trees usable if the live AND-entry limit is lowered

Testing

• cargo fmt --all --check
• cargo check --workspace --all-targets --locked
• cargo clippy --workspace --all-targets --locked -- -D warnings
• cargo test -p wok-query --locked
• cargo test -p wok-compat --test nip_conformance --locked (27 passed)
• cargo test -p wok-compat --test e2e_transports --locked -- --test-threads=1 (12 passed)
• workspace tests excluding wok-bench passed; one parallel LMDB allocation failure was confirmed transient by the serial E2E run

Includes unit, property, historical scan, live monitor, exact COUNT, router/sync fallback, negentropy, conformance, E2E, and fuzz coverage.

[erskingardner]

The NIP-91 matcher, overlap normalization, forced full-event verification, HLL omission, and limit plumbing look correct and well covered. Requesting changes for a compile break in the separate fuzz workspace; the other notes are interop/cost follow-ups.

Policy question, not a code defect: this advertises draft NIP-91 from nips#2252, which is not in the conformance pin 656cecc7…. That is a deliberate exception to “canonical NIPs it advertises.” Fine if maintainers want it, but it should be an explicit call rather than landing as if 91 were already in the pin.

[hzrd149]

Pushed follow-up commit 50227d2 addressing the review feedback:

• added the missing wok-query dependency to the standalone fuzz workspace
• aligned query admission cost with the tag/AND index actually selected
• select the smallest AND tag set as the historical scan seed
• added capability-aware forwarding for router REQs: fetch/cache upstream NIP-11 and preserve &* only when NIP 91 is explicitly advertised; otherwise send the # compatibility superset and retain exact local post-filtering
• applied the same NIP-91 filter handling to sync NEG-OPEN
• require upstream NIP 77 advertisement before sending NEG-OPEN
• coalesce concurrent NIP-11 lookups per normalized relay URL, with bounded process-local caching (1h valid results, 5m failures, 256 relays)
• retained Wok conditional NIP-91 advertisement when AND filtering is enabled

Validation completed:

• cargo fmt --all --check
• cargo check --workspace --all-targets --locked
• cargo test -p wok-query --locked
• relay capability tests
• 9 mesh/NIP-11 cache and forwarding tests
• cargo check --manifest-path fuzz/Cargo.toml --bin ingress --locked

The policy question about advertising draft NIP-91 remains an explicit maintainer decision; the implementation continues the PR behavior and advertises it only when max_tags_per_filter > 0 and max_and_entries > 0.

[erskingardner]

Previous findings are fixed: fuzz depends on wok-query, query cost follows the tag/AND index, AND seed picks the smallest AND set, and outbound filters are capability-aware.

The NIP-91 matcher is still in good shape, and advertising draft 91 is an accepted maintainer call.

Requesting changes for two follow-up issues in the new NIP-11 forwarding path. After those, this looks mergeable (CI has not reported on 50227d2 yet).

[erskingardner]

Previous request-changes items are fixed on f17c044.

• Compatibility path folds each &x set into #x instead of dropping keys, so an &-only filter cannot become {} / a remote firehose. Malformed & clauses still fail before the NIP-11 lookup, and filter arrays are rewritten element-wise.
• NIP-11 is used only to decide NIP-91; NEG-OPEN is left to prove 77. Unreachable or malformed documents degrade to the compatibility filter.
• Historical scans seed one cursor from the smallest AND set; estimated_cost charges that seed. nip91_and_tag_seeds_a_single_cursor covers the cursor-count regression.

The matcher, overlap normalization, forced full-event verification, HLL omission, restricted COUNT scoping, live monitors, search does_match_without_search, and limit plumbing still look correct. Draft NIP-91 advertisement remains the accepted maintainer exception, gated on both tag-filter limits being nonzero.

GitHub has not reported checks on this head; I reviewed the diff and tests as written, not a CI run of f17c044.