[squid-protocol]

Part of #4269.

Slice chosen by demand

Census (cics_census.py blockers --baseline --all-programs, origin/main d3b7440, 179 programs over burned + non-burned corpora): of all #4269 forms, only ROLLBACK reaches a translatable program -- WHENEVER, dynamic SQL, CALL, CONNECT, rowset FETCH: 0 programs; SCROLL: 1 (GenApp, not a gap). The only EXEC SQL gaps were ROLLBACK: the generated repositories commit each statement (3 programs, 2 one gap-class from whole) and no generated Db2 repository method (the same 3; 2 one-away) -- nexusbank NEND-DAY, NINT-CALC (batch) and NBLK-ACCT (CICS), all non-burned. The second turned out to be a det translator bug: the generated Javadoc's program name NEND-DAY did not match \w+. Once those were fixed, NINT-CALC's last gap was ROLLBACK TO SAVEPOINT, so savepoints are in this slice too.

Census before -> after (programs translated whole): 95 -> 96 / 179 (non-burned 9 -> 10). NINT-CALC is now whole (89 -> 97/97 statements). NEND-DAY goes 86 -> 100/101: its one gap left is a host variable the translator does not find, not SQL. NBLK-ACCT goes 118 -> 128/166, and its ROLLBACKs are now refused by name (CICS). EXEC SQL gap classes over all programs: 2 classes in 3 programs -> 1 (CICS ROLLBACK, 1 program).

What changed

• det translator (det/sql.py)
  • COMMIT [WORK] -> DetSql.commit, ROLLBACK [WORK] -> DetSql.rollback. SAVEPOINT ... ON ROLLBACK RETAIN CURSORS, ROLLBACK [WORK] TO SAVEPOINT [x] and RELEASE [TO] SAVEPOINT x -> DetSql.savepoint (run as written).
  • A cursor declared WITH HOLD opens with hold. A hyphenated program name now finds its generated methods.
  • Refused by name: a CICS program's ROLLBACK (z/OS answers -926; SYNCPOINT ROLLBACK is the task's), a savepoint statement in CICS or with a host variable, and COMMIT / ROLLBACK with other options. A CICS COMMIT keeps its old behaviour.
• det runtime (DetSql)
  • UnitOfWork, given by the runner, holds a batch step's Db2 transaction. COMMIT closes the cursors not declared WITH HOLD. A held cursor keeps no current row, so a positioned statement before the next FETCH gets -508.
  • ROLLBACK closes every cursor. With no unit of work given, ROLLBACK and savepoint statements are refused at run time and never answered.
• oracle (ggsql.c, equivalence_sql.py, ggabend.c)
  • COMMIT closes the cursors that are not held and ROLLBACK closes all of them, as z/OS does. CLI keeps cursors open by default, so ggsql.c closes and frees the handles itself; the precompiler marks a held OPEN with H.
  • CEE3ABD backs out the unit of work before exit, as z/OS does for a batch program that abends.
  • Two CLI quirks were measured and neutralised. CLI runs ROLLBACK WORK TO SAVEPOINT as a full ROLLBACK (0 where Db2 answers -880), so the precompiler drops the noise word WORK. CLI's row count of -1 no longer reaches SQLERRD(3).
• harness
  • A Db2 batch step runs as one transaction and gives the det runtime its connection, by reflection.
  • After a matching abend the Db2 tables are compared too: the abend backed out the uncommitted work, so they hold what the step committed.
  • The Equivalence harness: DB2 SQL support #4507 outcome fold no longer counts the stub's -501 on a cursor that COMMIT / ROLLBACK closed.
  • New equivalence.py run --corpus-dir, never with --record, for a synthetic estate.

IBM references (register Q3 / Q4, docs/language_status/oracle_assumptions.md)

Db2 for z/OS SQL Reference: COMMIT (closes cursors not WITH HOLD), ROLLBACK (backs out, closes all cursors), SAVEPOINT / ROLLBACK TO SAVEPOINT / RELEASE SAVEPOINT (-880, -881). Codes: -501, -508, -803, -925 / -926 (2D521, COMMIT / ROLLBACK invalid in CICS). Application Programming and SQL Guide: unit of work in TSO / CAF (commit at a normal end, backout at an abnormal one). Q3 goes from DIFFERS (batch) to MATCHED.

Oracle / z/OS differences, declared:

• A CICS program's EXEC SQL COMMIT is still a reset on the Java side and a real commit on the oracle, where z/OS answers -925. No burned CICS program issues one (CBSA's COMMIT is in BANKDATA, a batch program). This is the next slice.
• A libcob run-time failure that is not CEE3ABD still commits at exit. It never compares equal anyway: the Java side reports an UNCODED abend.

Proof

• Hand-traced, both sides, real Db2: tests/equivalence/db2/uow (synthetic UOWDEMO estate) + test_sql_unit_of_work.py::test_uowdemo_det_port_proves_on_db2_against_the_hand_trace, gated on EQUIVALENCE_E2E=1 and passing locally in normal and abend modes.
  • Covers a commit, two changes backed out, -803 then ROLLBACK, -501 after a COMMIT on a cursor that is not held, a held cursor kept by COMMIT, ROLLBACK closing it, a savepoint (insert backed out; -880 after RELEASE), and uncommitted work kept at a normal end and backed out by an abend.
  • SYSOUT (24 lines), the table and 18/18 statement outcomes match on both sides and match the hand trace.
  • No burned program issues ROLLBACK, and the census repos are blind, hence the synthetic estate.
• Negative control: with the Java ROLLBACK switched off, the e2e test fails.
• Unit tests (no Db2): the translator's output and refusals, DetSql with a fake unit of work run on javac/java, the precompiler's hold mark, and the harness's step wrapper.
• det_port.py check vs origin/main: 29 ports changed. Every Db2 port carries DetSql.java, and LGUPDB01 also gains the WITH HOLD flag. Re-proven with proof_sweep.py --det-only --cases <29> --faults all: 29/29 proven, "sweep: as expected".
• Gates: pr_gates.py --fast 5/5 pass; pr_gates.py --ratchets 7/7 pass, none skipped.
• Evidence: no committed port, case, corpus pin or declared difference changed. The harness and oracle changes make records stale only for the scheduled refresh. The evidence report is not committed.

Next slices (#4269)

1. EXEC SQL COMMIT / ROLLBACK in CICS: -925 / -926 on both sides. ggsql.c needs to know it runs under CICS (ggcics.c, CICS files).
2. NEND-DAY's last gap: a host variable the translator does not find (EXEC ... HV-SO-IM: no such item, line 408). Probably not SQL; it makes the program whole.
3. Savepoints in a CICS task, through the task's unit of work.
4. The forms with no demand in the census (0 programs): WHENEVER, EXECUTE IMMEDIATE / PREPARE / EXECUTE / DESCRIBE, CALL, CONNECT, rowset FETCH, SCROLL (1 GenApp declaration), DSNTIAR (Q5).

🤖 Generated with Claude Code

https://claude.ai/code/session_01MQNoe4wPJgr7dbUvs3DG6D

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab