[kvlonge]

D150/D320 regression test. Three corpora, each checked by analysis::tests::test_everyday_corpus under the built-in policy and the repository policy (UTC session).

Changes since 521b883 (the approved head, auditor record 822f907), in b2da27d and 0076e8e:

1. The provenance strings in both snowflake-receipts.json files now describe the schema neutrally, naming the corpus schema files. The verdicts are unchanged.
2. D334, the coordinator's ruling, makes the test asymmetric:
   • It fails on regressions: a statement that translated now refuses, a refusal changes its code or span, a check or parse diagnostic changes, a crash, or a translated basics statement returns a different result on corpora/basics/fixture.sql (sorted-row hash, or the Snowflake errno it raises).
   • It reports improvements without failing: only a statement recorded as refused … that now translates, or one recorded as unrecorded. A recorded parse or check diagnostic that disappears fails, because Snowflake may reject that SQL (auditor record 976fb43).
   • make corpus-rerecord rewrites every recorded outcome and result and keeps the files' comments. The failure message names this command, so a PR that moves an outcome on purpose runs it and commits the corpora in the same commit.
   • The basics generator now writes statements with an unrecorded outcome, and make corpus-rerecord fills it in. Its Python outcome recorder is removed, so the Rust test is the only recorder.

No converter code changes.

Since 0076e8e (record b64ede2), in a9e5755 and f1944aa:

• Main fc15d51 (D316 + D322) is merged in, and the corpora are re-recorded on it: 29 more queries translate under the built-in default (clock reads, unread VARIANT columns), and 20 refusals change code. No query regressed.
• D343: the basics generator's FIXED list gains SELECT order_id FROM orders WHERE ordered_at >= DATEADD(day, -30, CURRENT_DATE) and the same over CURRENT_TIMESTAMP(), as 10 entries with their spellings. Both are pinned as refused under both policies: SQ0003 node:gte and SQ0003 fn:DATEADD. They should flip to translated once fix: compare naive timestamps with CURRENT_TIMESTAMP() and count sub-day DATEDIFF of a DATE #115, C's group 1 and D326 land, and each of those PRs re-derives with make corpus-rerecord. No other basics entry changed.

At the slot (247ca79, 99e4ab3): main c2cbe9d merged in, then make corpus-rerecord. Only basics moves (D327): 387 refused → translated, 189 refusals change code, and 8 unaliased literals with a trailing comment (select 1 -- trailing comment and similar) go translated → refused with SQ0009, as D327 refuses them. D315 is expected to convert those again. No fixture result changed.

Corpora:

• corpora/everyday: 615 analyst queries, 537 of them valid per the receipts.
• corpora/everyday-models: 499 dbt-style models.
• corpora/basics: 1,842 generated basics with surface variation, over the playground's sample schema, with fixture rows.

Recorded translation rates on fc15d51, built-in / repository: everyday 376 / 459 of 537, models 293 / 404 of 458, basics 1,095 / 1,160 of 1,842. These pin SQLQuack's own outcomes. They are not Snowflake-verified support figures: only 48 of the 1,114 hand-written queries have receipts, and none of the basics do.

Checks:

• The three tests pass, in about 60 s, mostly the basics fixture runs.
• Re-recording is idempotent: the generator followed by make corpus-rerecord reproduces the committed basics file byte for byte, and a second re-record changes no corpus.
• Asymmetry, checked by hand-edited expectations and then reverted:
  • An expectation hand-flipped from translated to refused is only reported: the statement now looks like a refusal that translates.
  • Expectations hand-flipped to parse SQ0001 and to check SQ0004 on translating statements both fail.
  • A changed result hash fails, with the re-record hint.
  • A refused → translated expectation flip fails.
• clippy -D warnings, cargo fmt, ruff, ruff format and fensu are clean, and the hygiene scan passes.

The full D94 gate runs when the slot comes: merge main, make corpus-rerecord, then gate that exact head.

[kvlonge]

D94 gate passed on the exact head 99e4ab3: #135 at f1944aa (auditor record db54f0b), merged with main c2cbe9d, then make corpus-rerecord.

• make check-merge RECORDER=0: exit 0 in 41m36s. fensu reported 0 faults, and the three corpus tests passed.
• make verify-recorded: exit 0 in 22m53s. 12,033 cases: 7,997 passed, 1,035 diverged as expected, 2,964 refused, 37 typed, 0 pending, 0 unproven, 0 failed; 226 of 226 equivalences and rules verified. The translation results have 0 regressions against the baseline.
• The public-repository hygiene scan is clean.

GitHub Actions can't run because of the account billing block, so this local gate is the check for this landing.