[bpamiri]

Fixes #3715. Based on develop at b313a44, which includes #3710, so the BoxLang + Oracle leg gets past fixture population.

Root cause

The failure reproduces locally on BoxLang + Oracle (isolated compose project) with the same 4 insertAll errors as CI. Probes then isolated the cause with plain queryExecute calls, no Wheels code involved:

statement	outside a transaction	inside a transaction
INSERT ... SELECT ... UNION ALL	ORA-17009: Closed statement	Error closing connection: Connection is closed
INSERT ... SELECT (single row)	ORA-17009	Connection is closed
INSERT ... VALUES, UPDATE, SELECT	ok	ok
leading /* */ comment, WITH clause	ORA-17009	—
BEGIN INSERT ... SELECT ...; END; (with bind params)	ok (2 rows)	ok

BoxLang requests generated keys on every INSERT, whether or not a result is captured. Oracle's driver cannot return keys for INSERT ... SELECT, which is the shape $bulkInsertSQL must emit (#3302 per-row identity). In a transaction, the resulting error is masked by the connection-close failure. This is the same class of bug as #2745 and #3653.

Change

• OracleModel.$bulkInsertSQL wraps the statement: BEGIN INSERT ... SELECT ...; END;. It is still one statement with one source row per record, and all values are still bound. OracleMigrator already runs BEGIN ... END; blocks on every engine.
• oracleBulkInsertSqlSpec pins the block shape. Existing assertions are unchanged.
• The fix exposed one more masked error: bulkOperationsSpec measured Len() of a raw oracle.sql.TIMESTAMP (Cannot determine length of object). It now asserts that the column normalizes to a date via $normalizeDbTimestamp(), a stronger check.
• Changelog fragment added.

Verification

• BoxLang + Oracle, local, wheels.tests.specs.model: before, 1047 pass / 4 error; after, 1052 pass / 0 fail / 0 error / 14 skip.
• BoxLang + SQLite model: 1051 / 0 / 0.
• Lucee 7 + SQLite model (tools/test-local.sh): 1051 passed.
• Lucee / Adobe + Oracle: compat-matrix dispatch on this branch; totals to follow.

🤖 Generated with Claude Code

[bpamiri]

Review: independent reviewer (rev1.r1), commit 77ce2cb6b. Verdict: approve. No MUST-FIX. Merge gate, pending when I posted: dispatch 36344707369 (on this exact head). Read the result JSON, not the job colour, because Oracle is soft-fail.

• lucee6, lucee7 and adobe2023 (and adobe2025, if it gets past ci: Adobe 2025 + Oracle exhausts Java heap before producing test results #3709) + oracle: bulkOperationsSpec and oracleBulkInsertSqlSpec must pass. This change alters the Oracle bulk-insert SQL on every engine, not just BoxLang, so these legs are the real regression check for the anonymous-block form with bind parameters.
• boxlang + oracle: the 4 insertAll errors from fix(oracle): insertAll inside a transaction fails with "Connection is closed" on BoxLang #3715 should be gone. One TimestampRoundTripSpec error is expected there, because this branch has fix(oracle): read a generated-key object's text before string checks #3710 but not test(database): build the timestamp round-trip diagnostic without casts #3716.

Root cause and fix. BoxLang requests generated keys on every INSERT, and the Oracle driver can't return them for INSERT … SELECT … FROM dual, hence ORA-17009. Wrapping the statement as BEGIN INSERT … ; END; means the engine no longer sees an INSERT, so it makes no key request. The per-row identity shape and bind parameters are unchanged, and OracleMigrator already relies on anonymous blocks.

Checked for side effects:

• Return value. A JDBC update count for an anonymous block isn't the row count, so I checked where insertAll()'s number comes from. vendor/wheels/model/bulk.cfm computes insertedCount from the batch bounds (batchEnd - batchStart + 1), never from the driver, and doesn't read the result. So wrapping the statement can't change what callers get back.
• Size. Batches stay at 1000 rows, so the bind count per statement is unchanged; the block adds only BEGIN … END;.
• Transactions. The block has no COMMIT, so it takes part in the surrounding transaction as the bare INSERT did. Errors still surface, now as ORA errors with ORA-06512 context.
• The spec pins ^BEGIN INSERT INTO "AUTHORS", FROM dual; END;$ and exactly one BEGIN. The bulkOperationsSpec change from Len() to IsDate($normalizeDbTimestamp(...)) is stricter than before and no longer measures a raw driver object. It's the same masked-diagnostic pattern as fix(oracle): BoxLang reads a cf_sql_timestamp back as an unreadable oracle.sql.TIMESTAMP #3714.

[bpamiri]

Independent review of 77ce2cb: source review clean; Oracle matrix evidence pending.

Independently ran bash tools/test-local.sh model on Lucee 7 + SQLite: 1051 passed, zero failures/errors. Reviewed the full bulk-insert caller and SQL builder: the wrapper preserves ordered bind objects, UNION ALL row shape, identifier quoting, and batch boundaries; it adds no commit/rollback or exception handler. insertAll() computes its inserted count and does not consume a driver-generated key/result. OracleMigrator already uses anonymous blocks.

The timestamp assertion now verifies a normalized date instead of merely nonempty text; it retains the requirement that timestamps were populated.

Final clearance awaits actual Oracle results from dispatch 36344707369, whose head matches this PR. The change applies to all Oracle engines, so the BoxLang reproduction fix also needs Lucee/Adobe execution evidence. A soft-fail job colour alone will not suffice.

[bpamiri]

APPROVED at 77ce2cb.

Independently downloaded the actual Oracle JSON artifacts from exact-head dispatch 36344707369. All insertAll and upsertAll specs passed on BoxLang, Lucee 6/7, and Adobe 2023. Lucee 6/7 and Adobe 2023 each report 5,792 pass / 0 fail / 0 error / 24 skip. BoxLang reports 5,791 pass / 0 fail / 1 error / 24 skip; its sole error is the separate TimestampRoundTripSpec diagnostic issue #3714, addressed by #3716. Adobe 2025 remains pending and is not claimed as verified here.

The Oracle anonymous block keeps bound parameters, batching and transaction ownership intact; it does not commit or suppress errors. The caller already computes inserted count from the batch and does not consume generated keys. Earlier independent Lucee 7+SQLite model suite: 1,051 passed, zero failures/errors. No blocking finding in #3715's scope.

[bpamiri]

Follow-up from rev1.r1, for the record: the gate is met. I read the result JSON of dispatch 36344707369 (77ce2cb6b) myself for two legs, and dev1-design reported the rest:

• adobe2023 + oracle and lucee7 + oracle (my read): 5816 specs, 5792 pass, 0 fail, 0 error. bulkOperationsSpec 13/13 and oracleBulkInsertSqlSpec 6/6 pass, including wraps the INSERT … SELECT in an anonymous PL/SQL block. So the BEGIN … END wrapper with bind parameters is fine on Lucee and Adobe, which was the cross-engine risk.
• lucee6 + oracle is the same (via dev1-design). boxlang + oracle: all insertAll cases pass; its 1 error is TimestampRoundTripSpec (fix(oracle): BoxLang reads a cf_sql_timestamp back as an unreadable oracle.sql.TIMESTAMP #3714), fixed by test(database): build the timestamp round-trip diagnostic without casts #3716 and not on this branch. adobe2025 + oracle produced no results (ci: Adobe 2025 + Oracle exhausts Java heap before producing test results #3709/fix(dispatch): first requests on a cold Adobe 2025 start fail before Wheels finishes initializing #3726).
• Independently, fix(model): return CFML dates for Oracle datetime columns on every engine #3725's run on develop with both test(database): build the timestamp round-trip diagnostic without casts #3716 and fix(oracle): run bulk inserts as a PL/SQL block so BoxLang asks no keys #3718 gives a fully clean boxlang + oracle (5802 pass, 0 fail, 0 error).