[kvlonge]

D250: three launch blockers from corpus round 3 (dbt_utils.date_spine, dbt_date.get_date_dimension, FLATTEN over VARIANT), each converted from Snowflake receipts. A wrong answer TO_CHAR gave for MMMM is fixed too.

What changes

• POWER(2, n) of a whole NUMBER exponent converts as pow(2, n). A double holds every whole power of two from 2^-1074 to 2^1023 exactly, so there is no rounding question, unlike other bases. Fractional exponents and other bases still refuse.
• TO_CHAR MON and MMMM write the English month abbreviation and full name in title case (Jan, January). They convert only when written in upper case, as recorded; mon or Mmmm refuse.
• Fixes a wrong answer: on main, TO_CHAR(d, 'MMMM') converts to strftime(d, '%m%m'), giving '0909' instead of 'September', because the shorter MM matched both halves. The tokenizer now refuses any format where two adjacent fields spell one longer run of a letter (MMMMMM, DDDD, SSSS, YYYYYYYY).
  • A new test walks every documented format element (YYYY … UUUU, HH12, AM/PM, A/P, FF0–FF9, TZH…). Each converts as recorded or refuses.
  • TO_DATE matches whole recorded formats, and TO_TIMESTAMP's reader rejects a repeated element, so neither had the double match.
• LATERAL FLATTEN with PATH of literal dot-separated plain keys flattens the value at that path:
  • PATH is prefixed with the path: x.y[0], x.y.b, and x.y['a c'] for a key that isn't a plain name.
  • A missing path, a scalar, a JSON null or an empty array at the path gives no rows, and keys match by case.
  • PATH with RECURSIVE or OUTER, and other path spellings, still refuse.

Evidence

• Receipts, recorded by lane A:
  • evidence/d250-launch-blockers d66ba9a: POWER(2, n) for n from -52 to 64, 2^1023, 2^-1074, 2^-1075 = 0, 2^1024 = inf; MON and MMMM for every month; REGEXP_SUBSTR of an empty-matching pattern past the end, for Convert REGEXP_SUBSTR with a computed position, occurrence or group #112.
  • evidence/d250-flatten-path da60f28: PATH reaching an object, keys by case, and a scalar or null at the path.
• Replay: pow.power-of-two (5), to-char.month-names (3 of 4; the fourth refuses on the ',' separator), flatten.path (3), flatten.toml t002_path and flatten_recursive flatten_path_option all pass. The older DD-MON-YYYY cases (to-char t002_date_format) now pass too. 0 failed.
• The dbt repros convert and run on DuckDB: 3 * power(2, 1) gives 6.0, and lateral flatten(input => e.payload, path => 'items') gives the item rows.
• Full verify: 11,607 cases, 7,650 passed, 0 failed, 220 of 220 verified. Translation baseline against main e9f002d: 0 regressions and 17 changes. 15 statements now match, and 2 new statements refuse: month-names t003 on its ',' separator, and the REGEXP_SUBSTR probe, which Convert REGEXP_SUBSTR with a computed position, occurrence or group #112's computed rule handles.
• The new format-walk test passes. On main it would fail at MMMM and DDDD.

The D94 gate runs on the landing head when it's this PR's turn.