[magnusp]

Fixes #108

Contract

Hrana 3 is the source of truth for transport, transaction and stream state; java.sql is the source of truth for the JDBC surface; xerial/sqlite-jdbc is a behavioral reference, not a copy target.

What changed

setAutoCommit(false) already sends nothing under the lazy begin from #99, and beginPending is derived as !autoCommit && baton == null, so a failed BEGIN needs no separate state fix. This PR adds the missing state-pin case instead of production code:

• aFailedBeginPinsTheLazyBeginStateUntilTheRetrySucceeds fails the first BEGIN during a statement and then asserts:
  • getAutoCommit() stays false,
  • the failed BEGIN closes its stream, so no phantom transaction is left,
  • commit() and rollback() with no work send nothing and succeed,
  • the next statement retries BEGIN with a null baton (new stream) and then runs on the stream that BEGIN opened.

Production code is untouched: the derived state does not drift.

Evidence

Red (pre-#99 commit 8c9381b, where setAutoCommit(false) began eagerly), /tmp/xerial-parity-i108/red-pre99.log:

[ERROR] LibsqlConnectionTest.aFailedBeginPinsTheLazyBeginStateUntilTheRetrySucceeds:148 » SQL transaction end failed on server
	at LibsqlConnection.beginTransaction(LibsqlConnection.java:223)
	at LibsqlConnection.setAutoCommit(LibsqlConnection.java:300)

Green at the PR head (4811c7d), /tmp/xerial-parity-i108/lane-3-specific.log:

Tests run: 1, Failures: 0, Errors: 0, Skipped: 0
BUILD SUCCESS

Lanes (all commands with bare ./mvnw, as the queue brief requires instead of the plan's mise exec -- prefix):

Lane	Command	Result
1	new cases at trunk (1bc66f9) and head	pass at both; red at pre-#99 for the asserted reason (lane-1-regression.log, red-pre99.log)
2	commit/rollback with a pending begin	pass, sends nothing (lane-2-specific.log)
3	statement after a failed BEGIN	pass, retries BEGIN on a new stream (lane-3-specific.log)
4	-pl libsql-jdbc test	265 tests, BUILD SUCCESS (lane-4-module-suite.log)
5	-pl libsql-conformance -am verify	123 tests, BUILD SUCCESS (lane-5-conformance.log)
6	-Pxerial-control verify	100 tests, BUILD SUCCESS (lane-6-control.log)
7	spotless:check	BUILD SUCCESS (lane-7-spotless.log)
8	./mvnw test	BUILD SUCCESS (lane-8-reactor.log)
9	-Dtest=LibsqlConnectionTest	28 tests, BUILD SUCCESS (lane-9-specific.log)
10	-pl libsql-jdbc -am package -DskipTests	BUILD SUCCESS, jar built (lane-10-package.log)

Perf probe (perf.log), three interleaved runs of ./mvnw -pl libsql-jdbc test:

• trunk: 2166, 2115, 2102 ms → median 2115 ms
• head: 3450, 2110, 2094 ms → median 2110 ms

Head median is within the ten percent rule. The first head run is a compile/warm-up outlier after the package lane; the remaining two pairs differ by under 20 ms.

Divergences

• No production change. The plan allows this: the Non-auto-commit connections hold an idle Hrana stream that sqld expires #99 derived beginPending state already satisfies the pinned behavior.
• Lanes run as ./mvnw without the mise exec -- prefix, per the queue brief (mise attempts an unrelated tool install here).
• New test method name differs from the plan's wording; it covers the plan's failing-BEGIN-then-successful-retry plus the no-op commit and rollback checks.