docs(lance-backend): upstream test aggregation — 42/42 across 8 CRUD binaries under SURREAL_TEST_KV=lance

[AdaWorldAPI]

Summary

Adds .claude/lance-backend/UPSTREAM_TEST_RESULTS.md aggregating pass-rate for 8 upstream surrealdb integration test binaries under SURREAL_TEST_KV=lance. Verifies the kv-lance backend (Sprints A–W) end-to-end against real SurrealQL workloads, not synthetic unit tests.

Results

Test binary	Pass / Total	Wall time	Source sprint
create	3 / 3	49.8s	Sprint O (#1)
update	2 / 2	36.5s	Sprint S (#5)
select	9 / 9	167.8s	this sprint (pre-batch)
delete	3 / 3	37.7s	this sprint
insert	16 / 16	47.1s	this sprint
upsert	2 / 2	50.1s	this sprint
merge	1 / 1	0.6s	this sprint
relate	6 / 6	1.8s	this sprint
total	42 / 42

Zero failures. Zero compile errors.

What this verifies

Every test in the table exercises the full kv-lance path:

SurrealQL parser
    → planner
    → executor
    → Transactable::{get, set, put, putc, del, delc, scan, scanr, commit}
    → Lance MergeInsertBuilder (atomic upsert; Sprint N)
    → BTREE scalar index on `key` (Sprint M)
    → arrow-array 57 RecordBatch → disk via Lance v4.0
    → background optimizer compact_files + cleanup_old_versions (Sprint I)

The combined kv-lance + kv-mem feature set compiles because every tests/*.rs carries #[recursion_limit = "1024"] (Sprints O / S / V). The ndarray dep graph is clean (PR #10) — no [patch.crates-io], no ndarray-stats diamond.

What's NOT yet verified

22 integration test binaries still pending under SURREAL_TEST_KV=lance:

access, alter, asyncevent, auth_limit, cache, changefeeds, complex, define, field, function, future, index, info, live, param, query, remove, script, sequence, table, timeout, util

Each has the recursion_limit attribute already (compiles), so a follow-up sprint can batch them. The most-likely-to-surface-gaps candidates are define / query / index / function — covering schema definitions, complex queries, vector / fulltext indexes, and built-in function execution paths.

Why this matters

Up to this point, kv-lance verification was via:

• 50+ lance-specific unit tests (kvs::lance::tests)
• 3 SurrealQL smoke tests (kvs::lance::integration_tests)
• Property test against HashMap reference (Sprint J)
• Concurrent OCC tests (Sprint U)
• Earlier per-file integration verifications (Sprints O, S)

This sprint extends real-workload coverage to 8 CRUD-shaped binaries totaling 42 tests across creation / mutation / read / graph-edge operations — the surface every SurrealDB user touches.

Resolves the "Per-file pass rate aggregation under SURREAL_TEST_KV=lance" deferred item in .claude/lance-backend/KNOWN_DIFFERENCES.md.

Test plan

• CI doesn't need to re-run these (results are captured); the kv-lance backend has not changed since the runs above.
• Follow-up sprint will extend the table to cover the remaining 22 test binaries; failures there feed the next batch of compatibility-gap fixes.

---

Generated by Claude Code