mirror of
https://github.com/vitorpamplona/amethyst.git
synced 2026-10-05 19:28:25 +00:00
Claude
3b20aede9a
Adds a recorded corpus of 64 real events from search-staging.brainstorm.world — a vespa-relay, the same software this token language was ported from — and 18 tests over it: 8 on the engine in :commons, 10 on `LocalCache.filter` in :amethyst. Recorded, not fetched. `tools/search-parity/fetch_fixtures.py` drives `amy fetch` against the relay by hand; the tests read the committed fixture, so `./gradlew test` stays offline and the pre-push hook does not depend on somebody else's uptime. The fixture stores each case's filter FIELDS rather than a prebuilt filter, so the test rebuilds the Filter in view of the reader and a wrong rebuild cannot quietly make the assertions vacuous. What the relay can and cannot referee turned out to be the whole design, and it was measured rather than assumed: - **NIP-01 it can.** Given kinds, #t, since and until there is exactly one right answer, and across all 64 events our matcher agrees with the relay about every one it chose to return — 0 violations. That is now a hard assertion, with a converse test so it cannot pass by matching everything. - **NIP-50 it cannot.** This relay retrieves topically: asked for `bitcoin` it returns a block-height summary that never says "bitcoin". Eight of 64 events carry no literal occurrence of the term that fetched them. Asserting our substring matcher reproduces that would encode someone else's semantic expansion as a requirement on a lexical one — a test that fails on correct code. So text results are deliberately not compared, and the divergence is pinned as a range instead: zero would mean the relay turned lexical and the comparison should be rewritten, a quarter would mean we regressed. The fixture uses the `include:spam` lens, which waives the web-of-trust gate. Also measured: it makes the corpus reproducible, where `observer:<pubkey>` ties every answer to one account's moving trust graph — but it does not make retrieval lexical, and in fact widens the divergence from 4 events to 8 by letting more topical matches through. The LocalCache tests cover what the relay knows nothing about and where the bugs actually were: the regular/addressable split, the viewer-policy predicate composing with rather than replacing the filter, and the result cap keeping the newest — the ordering whose absence let `take(limit)` run before the sort. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017yKjw2WqwZpSzsqcYZMnkV