[kvlonge]

Unblocks C's slice 1 (land/d151-s1). Coordinator priority: this lands before the slice. It touches crates/sqlquack-syntax, so it goes on the auditor's D118 re-read list.

Bug. SQLQuack rewrote TRUNC/TRUNCATE's bare second argument into a date part ({"var": …}) whenever its spelling named a date part. TRUNC(x, s) with a NUMBER column s became "second". Snowflake recorded numeric truncation for those cases, and Polyglot parses s as a column. This was the only Rust failure slice 1 causes (runs::test_exact, 8 differences in trunc.trunc / trunc.truncate t001_int, t005_null_x, t006_null_scale, t007_scale_past). It was also the cause of trunc.truncate's SQ0002 node:var refusal.

Fix. The parser keeps the identifier as a column, as Polyglot does. Snowflake decides by scope, which the parser can't see:

• SELECT TRUNC ( 1 , DAY ) compiled with no table in scope (receipt sf.functions.trunc2.ka4427f986c78, accept, stage semantic).
• TRUNC(x, s) with a column s truncated numerically.

This reverses the TRUNC part of d704778, an auditor tree finding. DATEDIFF, TIMEDIFF, TIMESTAMPDIFF, EXTRACT and DATE_TRUNC keep their date-part rewrite. The test_fixes pins now assert a column for TRUNC(x, DAY) and TRUNCATE(x, s).

Zero-fire evidence

• Every recorded SQL text calling TRUNC or TRUNCATE on main was parsed with the old parser: 49 texts across cases/** and coverage/syntax/cases/*.jsonl. The rewrite fires on one, the trunc2 receipt above. Its verdict is unchanged, and sqlquack check still reports no problems for it.
• No translation case on main has a bare-identifier TRUNC unit, so no translation verdict moves.
• Syntax baseline, re-recorded with snapshot on the committed tree: the only change is sf.functions.trunc2 verified → unverified, because its auditor review fingerprints the old tree. Satisfied obligations (68,585), parsed and refused cases, and the 5,128 known receipt disagreements are identical. The auditor renews trunc2 under D118: 2,388 → 2,387 verified until then.
• sqlquack check: no problems for TRUNC(1, DAY), TRUNC(x, s) FROM t or TRUNC(d, day) FROM t.
• Polyglot directory check on those shapes: every entry identical, where all three differed before.

The converter is unchanged. A bare TRUNC unit was already refused (node:var), and a TRUNC scale from a column stays refused ("places other than a literal integer"). Resolving TRUNC(d, day) by type is a possible follow-up, which needs a receipt.