[kvlonge]

D228, part 2: verify replays a case's setup on DuckDB. Part 1 (the case format and recorder contract) made every setup case refuse. This replays the setup statements that have receipts and still refuses the rest.

What setup replays

Setup runs statement by statement, in order, and the whole case refuses at the first statement it can't replay:

• CREATE TABLE / INSERT INTO … VALUES: an unqualified, unquoted table with typed NUMBER and VARCHAR columns, created once, holding literals or NULL that each column holds without rounding. Relations resolve through {{run_database}} / {{run_schema}} and unqualified names.
• SET / UNSET of session variables (rule.session_variable): to a literal, NULL or a deterministic scalar subquery. $name reads it in either case. A variable setup didn't SET, an UNSET of one it never set, an expression, and a subquery whose result can change between calls (RANDOM, the clock) all refuse. DuckDB computes a subquery's value at each read, while Snowflake computes it once at SET, so a CREATE TABLE or INSERT while such a variable is still set refuses (review round 1).
• CREATE TYPE name AS data_type: a type spelled as that plain name, in either case, resolves to its base type in casts, typing and later setup columns. That's what Snowflake's receipts report: a::age is NUMBER(3,0), and a + b over two age columns is NUMBER(4,0). Quoted names, redefinitions, OR REPLACE, IF NOT EXISTS, qualified names and non-alias definitions refuse. So does a built-in type's name: one SQLQuack converts, or one that doesn't parse as a custom type, such as GEOGRAPHY (review round 1).

Not replayed (still refuse)

• CREATE SEMANTIC VIEW / SEMANTIC_VIEW(…): the parser keeps CREATE SEMANTIC VIEW only as raw text (node:raw). SEMANTIC_VIEW(…) stores METRICS and DIMENSIONS in separate fields, so it loses the clause order that decides column order (semantic-view t002). Both need parser work before a rule can be receipt-backed.
• ALTER SESSION and CREATE VIEW in setup: no case uses either, so no receipt backs a rule.
• A cast to a structured type (data-types-user-defined t003, ::address over OBJECT(…)) was already refused, and still is.

Brought in from land/d151-6n2b (ruling D306)

rule.session_variable needs a passing case, so this PR carries the session-variables cases and their receipts, byte for byte, and nothing else from that branch: no recorder, no other cases, no obligations.

Commit here	From	Paths
5e3a016	C's 487be66 (author and date kept)	cases/translation/expressions/session-variables.toml
7439a34	A's 7a9436d (author and date kept)	evidence/results/{13884601cc86…,24b0b834ddee…,bbd0b6d5c294…,c41db8d3b834…,fbd897b012ca…}.{json,arrow}, the receipts of t004, t003, t001, t005 and t002

All 11 blobs are identical to their source commits' (git rev-parse <commit>:<path>). The evidence guard passes, with 10 added result files. The five cases are uncited until C lands their obligations in obligations/expressions.toml, which are left out here. A and C have been told these paths land through this PR.

Evidence

In a scratch merge with land/d151-6n2b, the 21 setup cases give:

File	Passed	Refused	No receipt
data-types-user-defined	t001, t002, t004 (100046 overflow), t005	t003 (structured cast)
name-resolution	t001, t002, t003, t005		t004
session-variables	t001–t005
semantic-view		t001–t004, t006	t005

That's 13 passed and 0 failed. Unit tests in test_case_setup cover each accepted and refused shape, including:

• a quoted "age" that must not resolve, an alias of an alias, and a setup column typed by an alias
• a write after a subquery SET, which gives 2 instead of Snowflake's 1 without the round-1 fix
• the same write after an UNSET, which replays

SEMANTIC VIEW is a separate post-launch item (D306).

Gate

Gate (D94) on e56c888, which merges main de14f81 and is 98bdfd0 minus the baseline re-record:

• make check-merge exits 0: clippy, fensu (0 faults) and all test binaries pass.
• verify --require-verified --strict-errors: 12,012 cases, 7,858 passed, 1,032 diverged as expected, 3,085 refused, 37 typed, 0 pending, 0 unproven, 0 failed. 221 of 221 equivalences and rules are verified, including rule.session_variable.
• translation-results --against main de14f81: 0 regressions and 11 changes, all gains. sf.expressions.session-variables goes Unmeasured → Matching (59 → 60 of 1,183 forms), its five obligations are met, and its five statements match.
• 98bdfd0 only re-records coverage/translation/results-baseline.json from that run. It checks clean against itself.
• The session-variables obligations are C's 5840035, cherry-picked as 9d24a0d with an identical blob. The auditor approved them (record a8dc7de).
• Main has since moved to 78910c3 (D267: statistics aggregates, ADD_MONTHS, MONTHS_BETWEEN and NVL2 from receipts #127). This PR re-merges and re-gates when it gets its slot. The auditor noted a conflict with BINARY: the type, the VARCHAR→BINARY cast and byte-wise text functions (D284) #129 in catalogue/constants.rs and the baseline.

Draft, for the auditor. GitHub Actions is blocked by billing, so the local D94 gate counts.

[kvlonge]

Closing: SQLQuack lands through local-gate trains, not PR merges (D94), so open PRs only add clutter (D628/D630). The branch is kept. Any work here that isn't on main yet is tracked by the train owner and lands through a train.