test: address 3 remaining leangraph CI flakes (TableSearchTest, AvoidDuplicateLinkTest, SearchTest source-synced)

[sriram-atlan]

Summary

After MS-1267 (lean-graph fix), #2500 (InsightsTest retry threshold), and #2506 (SuggestionsTest awaitConsistency) all shipped, the daily Test (leangraph-test) workflow holds at 3 stable failures. Each is a real test-side issue with a distinct shape — none reflect a server-side bug.

Test	Failure	Fix
TableSearchTest.searchComplexTypes	expected [user1] but found [user2]	Index sourceReadRecentUserRecords by recordUser instead of asserting by list-position. ES doesn't guarantee struct-array element ordering across runs.
AvoidDuplicateLinkTest.idempotentAddName / .idempotentAddURL	expected [2] but found [1] / [3] but found [2]	Use the existing condition predicate of retrySearchUntil to also wait for links.size to reach the expected count. The Database row indexes synchronously, but Link → Database edges propagate separately.
SearchTest.findSourceSyncedAssets	assertFalse(tables.isEmpty()) returns empty	Data gap, not a test bug — confirmed via direct ES probes. Tenant has the development Snowflake connection but no tables under the expected schema and no entities tagged Confidential. Happens when perfect-demo loadTenant runs with --skip. Throw SkipException with a clear message rather than fail.

Diagnosis per case

1. TableSearchTest.searchComplexTypes (ordering flake)

// before — fails when ES returns user1/user2 in non-creation order
assertEquals(found.getSourceReadRecentUserRecords().get(0).getRecordUser(), "user1");
assertEquals(found.getSourceReadRecentUserRecords().get(1).getRecordUser(), "user2");

// after — order-insensitive lookup
Map<String, PopularityInsights> recentByUser = ...stream()
    .collect(Collectors.toMap(PopularityInsights::getRecordUser, r -> r));
assertEquals(recentByUser.keySet(), Set.of("user1", "user2"));
assertEquals(recentByUser.get("user1").getRecordQuery(), "query1");
// etc.

2. AvoidDuplicateLinkTest link-count race

PackageTest.retrySearchUntil already supports an optional condition: ((IndexSearchResponse) -> Boolean)? predicate (PackageTest.kt:241). Use it to also wait for the response's first Database's links.size to reach the expected count, before reading.

val response = retrySearchUntil(request, 1) { resp ->
    val first = resp.assets?.firstOrNull() as? Database
    (first?.links?.size ?: 0) >= expectedCount
}

Also factored the two near-identical validateTwoLinks / validateThreeLinks into one parameterized validateLinkCount(count, names, urls, captureGuid). Drops the redundant follow-up Database.get(..., true) refetch — the search response already populates links via includeOnRelations(Link.NAME, Link.LINK, Link.STATUS).

3. SearchTest.findSourceSyncedAssets (data gap)

Direct ES probes against leangraph-test.atlan.com:

1) "development" Snowflake connection: 1 result (qname=default/snowflake/1772405975) ✅
2) Tables under <conn>/ANALYTICS/WIDE_WORLD_IMPORTERS/CONFIDENTIAL: 0 results
3) Any tables tenant-wide with Confidential tag: 0 results

This isn't an SDK / lean-graph bug — the perfect-demo seed simply didn't run its Snowflake crawl + tag-attach step on this tenant (likely because loadTenant was invoked with --skip, which short-circuits the parallel-asset-load path). The test has nothing to validate against. SkipException keeps the test signal meaningful: failures = real regressions, not setup gaps.

Test plan

• CI build green on this PR
• After merge, Integration (TableSearchTest) and asset-import: chunk 0 are green on the daily Test (leangraph-test) workflow
• Integration (SearchTest) shows findSourceSyncedAssets SKIPPED (not failed) on tenants without the Snowflake-crawl data; remains green elsewhere
• Non-leangraph nightly Test workflow continues to pass (no regression on the JG-backed tenant where the Snowflake data IS present)

🤖 Generated with Claude Code