fix(pg/parser): KB-1/2/3 closure + Class A systematic silent-accept cleanup (#107)

Stacked on #105 #106. Closes oracle-discovered KB-1/KB-2/KB-3, audits + fixes 27+ Class A silent-accept sites, lands G1-GATE lint enforcement. See PR for full description.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Author: rebelice

pg/parser/PAREN_AUDIT_DRIFT_BASELINE.txt

@@ -46,6 +46,7 @@ publication.go	parseRuleActionMulti
 schema.go	parseAggregateWithArgtypesForComment
 schema.go	parseFunctionWithArgtypesForComment
 schema.go	parseOperatorWithArgtypesForComment
+schema.go	parseSchemaStmtCreateOptTemp
 select.go	parseFuncAliasClause
 select.go	parseOptAliasClause
 select.go	parseRelationOrFuncTable

pg/parser/PAREN_FOLLOWUPS.md

@@ -8,9 +8,9 @@ The following are post-merge items surfaced by Codex review at each phase. None
 
 Discovered by Phase 2/3 oracle + fuzz. See `PAREN_KNOWN_BUGS.md` for full entries.
 
-- **PAREN-KB-1** — `(T JOIN U)` accepted without ON/USING. Fix in `parseJoinedTable` / `parseJoinQual`. Pinned by §2.7 `t.Skip` + fuzz known-mismatches.
-- **PAREN-KB-2** — `Parse` returns only first RawStmt for multi-statement input like `SELECT * FROM (SELECT 1) SELECT 1`. Fix in `parser.go:Parse`. Top-level statement-list issue, out of `parenBeginsSubquery` scope.
-- **PAREN-KB-3** — `LATERAL ()` accepted with empty body; PG rejects. Fix in `parseSelectWithParens` nil-body check. Pinned by §3.2 `t.Skip`.
+- **PAREN-KB-1** — ~~`(T JOIN U)` accepted without ON/USING~~ **CLOSED 2026-04-22** (commit a593131) via `parseJoinQual` strictness fix. Unskipped §2.7 oracle test; fuzz known-mismatches allowlist now empty.
+- **PAREN-KB-3** — ~~`LATERAL ()` accepted with empty body~~ **CLOSED 2026-04-22** (commit 3eed1fe) via `parseLateralTableRef` nil-body guard. Unskipped §3.2 oracle test.
+- **PAREN-KB-2** — ~~`Parse` accepts multi-statement input without `;` separator~~ **CLOSED 2026-04-22** via 4 upstream-blocker fixes (KB-2a/b/c/d) + parser.go:Parse needSeparator check reland. See PAREN_KNOWN_BUGS.md for the full commit chain.
 
 ## CI hardening follow-ups
 

pg/parser/PAREN_KNOWN_BUGS.md

@@ -77,7 +77,72 @@ so the signal isn't lost when future work reads those files.
     `TestParenOracleLateral/invalid_shapes_rejected/LATERAL_empty_parens`
     (skipped with a pointer to PAREN-KB-3).
 
-## PAREN-KB-2 — `Parse()` silently drops trailing statements
+## PAREN-KB-2 — `Parse()` accepts statements without `;` separator — **CLOSED 2026-04-22**
+
+**Status:** All 13 upstream blockers fixed (KB-2a/b/c/d), then `parser.go:Parse` `needSeparator` check re-applied. pgregress + oracle suites fully green. The KB-2 entry below documents the history of the attempt, the blockers surfaced, and the fix commits that unblocked it.
+
+Closed by commit chain:
+- 0f5b7d2 KB-2a: parseRuleActionStmt NotifyStmt case
+- 6075da0 KB-2b: ALTER SEQUENCE SET LOGGED/UNLOGGED
+- 1808371 KB-2c: CREATE SCHEMA inline schema_element list
+- d7e9ba3 KB-2d: CREATE TEMP VIEW inside SCHEMA + DROP FUNCTION empty name
+- <this commit> KB-2 reland: needSeparator check in parser.go Parse loop
+
+## PAREN-KB-2 — `Parse()` accepts statements without `;` separator — **history**
+
+**Status:** fix attempted 2026-04-22 and reverted. Enforcing "cur must be `;` or EOF after each parseStmt" at parser.go:Parse surfaced 13 new pgregress failures — all pre-existing omni parser gaps that had been **masked** by the silent-accept behavior (CREATE RULE `DO INSTEAD NOTIFY x` body, SET inside transaction / savepoint blocks, CREATE-chained DDL). Before the statement-list strictness can land, those 13 upstream grammar gaps need to be fixed so the corresponding single-statement parses emit the right AST instead of truncating and letting a second parseStmt pick up the tail.
+
+### A1 KB-2 full-surface survey results (2026-04-22 re-run)
+
+Captured via clean baseline + in-place patch apply + clean revert, across all omni test surfaces:
+
+- `go test ./pg/... -count=1`: 13 NEW FAILURE entries (all in pg/pgregress).
+- `go test -tags=oracle ./pg/parser/... -count=1`: 0 additional NEW FAILURE entries (full 188-probe oracle corpus still green under the patch).
+- Other packages (pg, pg/ast, pg/catalog, pg/completion, pg/parsertest, pg/plpgsql/parser): all still green.
+
+**Total unique blockers: 13.** No new classes beyond what the delta reveals — oracle fence ran clean even under the stricter Parse loop.
+
+Concrete blocker list, grouped by underlying parser gap (earlier draft mis-attributed some files; this is the accurate breakdown from the 2026-04-22 re-run):
+
+**Category KB-2a — CREATE RULE + NotifyStmt action (4 failures, fix site: `publication.go:689 parseRuleActionStmt`)**
+- `copydml.sql:69` stmt[48] — `create rule qqq as on insert to copydml_test do instead notify copydml_test`
+- `rules.sql:1017` stmt[474] — `create rule r4 as on delete to rules_src do notify rules_src_deletion`
+- `with.sql:1721` stmt[291] — `CREATE OR REPLACE RULE y_rule AS ON INSERT TO y DO INSTEAD NOTIFY foo`
+- `with.sql:1726` stmt[293] — `CREATE OR REPLACE RULE y_rule AS ON INSERT TO y DO ALSO NOTIFY foo`
+
+Root cause: `parseRuleActionStmt` enumerates 4 of 5 PG `RuleActionStmt` alternatives (SELECT/INSERT/UPDATE/DELETE) — NotifyStmt falls through `default: return nil, nil`. The 5th production is literally named in the function's own doc comment but not handled.
+
+**Category KB-2b — ALTER SEQUENCE SET LOGGED/UNLOGGED (4 failures, fix site: ALTER SEQUENCE action parser)**
+- `identity.sql:535` stmt[266] — `ALTER SEQUENCE identity_dump_logged_a_seq SET UNLOGGED`
+- `identity.sql:537` stmt[268] — `ALTER SEQUENCE identity_dump_unlogged_a_seq SET LOGGED`
+- `sequence.sql:277` stmt[154] — `ALTER SEQUENCE sequence_test_unlogged SET LOGGED`
+- `sequence.sql:279` stmt[155] — `ALTER SEQUENCE sequence_test_unlogged SET UNLOGGED`
+
+Root cause: PG 15+ added `SET LOGGED` / `SET UNLOGGED` to the ALTER SEQUENCE action list. omni's ALTER SEQUENCE action dispatcher stops at an earlier match, leaves `SET` / `UNLOGGED` / `LOGGED` as residual tail.
+
+**Category KB-2c — CREATE SCHEMA schema_element inline body (3 failures, fix site: `parseCreateSchemaStmt`)**
+- `create_schema.sql:20` stmt[5] — `CREATE SCHEMA AUTHORIZATION regress_create_schema_role` + following inline `CREATE TABLE ...` etc.
+- `create_schema.sql:33` stmt[11] — `CREATE SCHEMA AUTHORIZATION CURRENT_ROLE` + schema_element list
+- `create_schema.sql:45` stmt[16] — `CREATE SCHEMA regress_schema_1 AUTHORIZATION CURRENT_ROLE` + schema_element list
+
+Root cause: PG CREATE SCHEMA grammar supports `CREATE SCHEMA [name] [AUTHORIZATION role] [schema_element [...]]` where schema_element is CREATE TABLE / CREATE VIEW / CREATE INDEX / etc. inline. omni's parser stops after AUTHORIZATION, leaves subsequent CREATE tokens as "next statement".
+
+**Category KB-2d — fringe / TBD (2 failures, need spot-check):**
+- `errors.sql:139` stmt[38] — trailing "CREATE" after some preceding stmt in the DROP AGGREGATE test context. SQL shown by regress harness is `-- should fail` (comment) but the real content is the preceding DROP AGGREGATE stmt whose body isn't fully consumed.
+- `create_view.sql:162` stmt[36] — trailing `(` inside the "-- subqueries" CREATE VIEW test block. Likely a CREATE VIEW parser gap when the view body has subqueries.
+
+Root cause: unknown without reading the extracted SQL precisely; spot-check at fix time.
+
+**Recommended path:**
+1. Open follow-up issues for each of the 3 pattern classes (CREATE RULE action body, transaction-block SET, CREATE TRIGGER body).
+2. Fix each so the single statement consumes all its tokens.
+3. Then re-apply the KB-2 fix — `needSeparator` flag in parser.go's Parse loop requiring `;` or EOF between parseStmt calls. The fix itself is ~5 lines and already designed in commit history (see reverted diff at pg/parser/parser.go Parse loop on 2026-04-22).
+
+Keep the oracle signal:
+- `pg/parser/testdata/paren-fuzz-corpus/seed-cases.txt` — NOT pinned (this is a parser.go-Parse bug, not a `(` dispatch bug).
+- `pg/parser/testdata/paren-fuzz-corpus/known-mismatches.txt` — NOT pinned; fuzz generator's `genReservedMisuse` no longer emits trailing-SELECT patterns.
+
+## PAREN-KB-2 — `Parse()` silently drops trailing statements (original entry, kept for history)
 
 - **Summary:** omni's top-level `Parse(sql string)` accepts inputs
   like `FROM (SELECT 1) SELECT 1` by returning only the first