[kvlonge]

Decision D267 (function rules from the website lane's corpus round 4): the statistics aggregates, ADD_MONTHS/MONTHS_BETWEEN and NVL2, each converted only as the recorded receipts show. Evidence comes from A's e510d1a (23 edge cases) and a586f83 (variance ties, scales, large-mean FLOAT).

Rules

Rule	What it converts	How the receipts pin it
rule.nvl2	NVL2(a, b, c), exactly three arguments, a not semi-structured	CASE WHEN a IS NOT NULL; branches unify without reading text, as recorded for NUMBER, FLOAT, text and untyped NULL. Text beside another type refuses.
rule.variance	VARIANCE, VAR_SAMP, VAR_POP, VARIANCE_POP, STDDEV, STDDEV_SAMP, STDDEV_POP of a NUMBER (scale ≤ 12), aggregate or window	The exact (n·Σx² − (Σx)²) / (n(n−1)), or / n², rounded half up to NUMBER(38, min(6+2s, 12)). The t009 tie 1/128 gives 0.007813, and the scales 1, 3, 5 and 12 are recorded. STDDEV is the FLOAT sqrt of that rounded variance, bit-exact on every receipt.
rule.variance_float	The same of FLOAT or text, without DISTINCT	The same formula in doubles; 6.583333333333333 is matched exactly, while the two-pass formula misses by 1 ulp. Sums go through the float_sum_order guard.
rule.covariance	COVAR_SAMP, COVAR_POP of NUMBER/FLOAT pairs, aggregate or window	(Σxy − Σx·Σy/n) / d over complete pairs, the only one of four candidate orderings that matches all 12 receipts exactly. Sums go through float_sum_order.
rule.add_months	ADD_MONTHS of a DATE, TIMESTAMP_NTZ, or TIMESTAMP_TZ whose offset is known	DATEADD by month, the count rounded half away from zero (2.50 → 3, −1.50 → −2); a month's last day stays a last day (2023-02-28 + 1 → 03-31), otherwise the day is clamped; the time of day is kept. The DATEADD core is shared, not copied, and DATEADD/TIMEADD verdicts are unchanged (98/8 before and after).
rule.months_between	MONTHS_BETWEEN of DATE/TIMESTAMP_NTZ	Whole months where the days match or both are month ends, with the clock ignored; otherwise plus (Δdays + Δclock) / 31 days, divided exactly (nanoseconds) and rounded at six decimals into NUMBER(27,6). Six-decimal ties are impossible for DATE inputs (31 is odd); no recording shows a timestamp tie.

What still refuses, on purpose

• CORR (all 12 cases): no formula reproduces the NUMBER(10,2) receipt 0.9974958437759694 exactly, so it is not approximated. Its obligations cite those refused cases, so CORR stays refused and fails closed.
• VARIANCE(DISTINCT) and any statistic DISTINCT over an ordered window: Snowflake rejects both (0A000 / 42601). DISTINCT is accepted elsewhere, as recorded.
• ADD_MONTHS of a TIMESTAMP_TZ column whose offset isn't known (t003).
• The SYSTEM$TYPEOF decimal cases (t007/t008): refused by the existing TYPEOF guard, not by these rules.

Declared divergences

• The large-mean FLOAT cases (variance t011, covar_samp/covar_pop t006) now declare float_sum_order. Their sums of products near 1e18 cannot be proven order-exact, so the converted SQL raises loudly instead of guessing Snowflake's cancellation.
• Policy: the statistics join numeric_range (a NUMBER total beyond 38 digits raises) and float_sum_order; ADD_MONTHS joins timestamp_range, as DATEADD already does.

Refactors (no behaviour change)

• conversion/constants.rs was at 2,002 lines after main merged in. The TO_CHAR format-element table moved into a text_formats domain together with strftime_pattern/clock_format, its readers. The statistics constants sit beside their only reader.

Measured

• Full verify on this head (current main merged in): 11,826 cases, 0 failed; 226/226 rules verified. Before this change the D267 cases refused; now +150 passed, +3 declared divergences.
• Forms, against current main's baseline: 0 regressions (89 → 98 matching without the D272 loader); the baseline is re-recorded. 9 forms go refused → matching: NVL2, MONTHS_BETWEEN, VAR_SAMP (aggregate and over), VARIANCE_POP, STDDEV_SAMP, STDDEV_POP, COVAR_SAMP over, COVAR_POP over. The rest of the family is held back only by the refusals listed above.

Draft. GitHub Actions is blocked by billing; the local D94 gate counts.

[kvlonge]

Landing gate (D301) on c70e255 (fd1c3b9 + main de14f81; the only conflict, results-baseline.json, taken from main):

• make check-ci: exit 0
• make test-rust: every suite passed
• make verify-recorded: 12033 cases, 7997 passed, 1035 diverged as expected, 2964 refused, 37 typed, 0 pending, 0 unproven, 0 failed; 226 of 226 rules.
• Translation results against main: 0 regressions, 299 other changes (this PR's rules moving statements from refused to matching, and its new statements). 78910c3 re-records the baseline from that run; against main it still shows 0 regressions, and the baseline holds the measure. Forms: 68 of 1183 match.
• Audit: approved at fd1c3b9; the merge adds only main.