Closes #1024.

Two unit tests flaked under full-suite CI load; both pass in isolation. Test-only fix — no product code changes (git diff touches only *_test.go).

Root causes

services/validator — TestValidateTransactionBatch_DuplicateOutpointCreatesConflicting (expected 36, got 4)
ValidateTransactionBatch validates the two txs concurrently; both spend the same outpoint through one shared sqlitememory store. The sqlitememory DSN (util/sql.go:218) sets no busy_timeout and leaves the pool unbounded, so under CI CPU starvation concurrent writes hit "database is locked"; the spend batch's 3 short retries exhaust → StorageError, and the validator collapses any non-UTXO spend error into ERR_PROCESSING (Validator.go:723) instead of the conflict path (ERR_TX_CONFLICTING). The #1017 batcher attribution in the issue is at most a contributing timing factor — the root vulnerability predates it.

services/legacy/netsync — TestSyncManager_createUtxos_ChunkFailureCancelsSiblings (4 not <= 1)
Racy by construction, unrelated to #1017 (no batcher on this path). errgroup cancels mergeCtx only after the failing worker returns, so the surviving worker keeps passing its loop-top mergeCtx.Err() check and draining iterations until cancellation propagates. The product short-circuit is correct; the test's "≤1 in flight" assumption is not.

Fixes (test-only)

• validator: utxoStore.RawDB().SetMaxOpenConns(1) — serialise the dev-mode SQLite pool so concurrent batch validations can't surface a transient lock. The concurrent batch path is still exercised; only DB access serialises. Production uses Aerospike.
• netsync: replace the 2-worker race with a deterministic single-worker test. The first SetMinedMulti cancels the parent ctx and returns success; because one worker runs sequentially, iteration 2's loop-top check always observes the cancel. Asserts zero post-trigger calls. Removing the short-circuit fails the count assertion.

Out of scope (flagged follow-ups, not bundled)

• Validator.go:723 collapses transient/storage spend errors into ERR_PROCESSING, discarding conflict semantics — worth a separate issue.
• sqlitememory has no busy_timeout and an unbounded pool; other concurrent SQL-backed tests share the same latent flakiness.

Test plan

• go test -race -count=50 each test individually — pass
• netsync mutation check: removing the mergeCtx.Err() short-circuit fails the test at the count assertion (expected 0, got 1)
• validator flake reproduced on base under -race -count=80 -cpu=8 (got 4); fix holds 360 iterations
• go test -race ./services/legacy/netsync/ ./services/validator/ — both green
• go vet clean