The watch-page/chart "exact" count relied on the NIP-45 COUNT, but
countEventsAcrossRelays returns bestEstimate = the single relay with the MOST
events — it can't union across relays. So when views are distributed across
relays it under-counted (a video's card showed 19 while the chart's uncapped
list union showed ~63).
exactCountForPointer now lists view events across ALL relays (one on-demand
list, not a storm) and folds the UNION into the deduped dedupeBuckets (one entry
per viewer-window). total = the set size, so it is accurate AND monotonic (never
spuriously decreases on partial relay responses). The per-relay COUNT is kept
only as a lower-bound floor when the list truncates (>2000 events), flagged
partial. The popularity modal passes { exact: true }; since the count flows
through shared pointer state, opening a video corrects its count everywhere.
Grid cards (no exact) still use the §17-capped batched backfill (approximate,
storm-safe) until opened — an accepted accuracy/cost tradeoff.
Also note the ⋯ menu "Event Details" button does nothing (new TODO #31).
Tests: testExactCountUnionsAcrossRelays. Grandfathered viewCounter.js size.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Triaged the monolithic tests/view-counter.test.mjs and got the whole file green,
then removed it from the QUARANTINE map:
- Pinned nostrClient.getSubscriptionManager() to null so the test exercises the
legacy per-pointer hydrate/subscribe path it mocks. The batched
SubscriptionManager path (added with the relay-storm work) was silently
shadowing the mocked legacy APIs.
- Made the dedupe-window test bucket-aligned (it used Date.now(), so the
+window/2 offset could straddle a fixed-bucket boundary and flake by time of
day).
- Corrected the hydrateVideoHistory root-timestamp assertions to the refactored
unit boundary: hydrateVideoHistory populates the rootCreatedAtByRoot MAP; the
video object's rootCreatedAt FIELD is applied by syncActiveVideoRootTimestamp.
reactionCounter.js audit (kind-7 likes): accurate, no fix needed. It does not use
NIP-45 COUNT; it lists kind-7 events and dedupes via applyReactionToState keyed
by pubkey (one reaction per user, latest-wins) with per-content totals — no
raw-count inflation. Kind-7 isn't replaceable, so listing+dedupe is the only
accurate model.
Spec corrections documented in TEST_INTEGRITY.md (SCN-view-counter-legacy-path,
SCN-hydrate-root-timestamp).
Co-Authored-By: Claude Opus 4.8 <[email protected]>
- Updated `hydratePointer` in `js/viewCounter.js` to only update local view counts if the network count is strictly greater (or if local state is uninitialized).
- This prevents regressions where a partial or stale network response would overwrite a higher, accurate local count.
- Updated `tests/view-counter.test.mjs` to reflect that fresh view events are additive to the cached high watermark (fixing a test expectation that relied on the old non-monotonic behavior).
- Updated `tests/view-counter.test.mjs` to match `resolveVideoPointer` behavior which prioritizes `dTag` (NIP-33) over `videoRootId`, fixing a legacy test failure.
- Removes the logic in `restoreCacheSnapshot` that deleted cached view counts if they were older than `VIEW_COUNT_CACHE_TTL_MS`.
- Updates `tests/view-counter.test.mjs` to assert that stale cache data is preserved and used as a baseline during hydration, preventing view counts from dropping to zero or lower fallback values during network issues.
