[kvlonge]

Decision D272 (with D272a/D272b): the translation figure is now bound to the translation obligations. Before this, nothing read coverage/translation/obligations/*.toml.

What changes

• syntax-coverage results loads every obligation file fail-closed. These are load errors that fail the command:
  • an unknown key
  • a form the catalogue does not hold, or that is absent from its source
  • a file named for a family that isn't the form's
  • an id that isn't <form>.t<NNN>, or that is declared twice
  • a boundary without a known subkind, or a subkind on another kind
  • a case outside cases/translation/<family>/…toml#, a case cited twice, or a case no recorded statement is
  • a required obligation with neither cases nor a non-empty held
  • a blank held or reason; held with cases or on required = false; reason on a required obligation
  • in retired-obligations.toml: a retired id that is reused or retired twice, an entry without a form and reason, or a replaced_by entry that isn't a live obligation
• An obligation is met only when it cites at least one case and every cited case is matching or matching_error. A cited case file must exist and its construct must be the obligation's form.
• A form is matching only if, in addition, it declares at least one required obligation and every one is met. Otherwise it gets one of the new states: unmet_obligations, or no_obligations when it declares no required obligation.
• Held (required = true, cases = [], held = "…") is never met and keeps its form out of matching. Unmeasured (required = false, cases = [], reason = "…") is never met and doesn't block its form.
• The JSON report gains:
  • obligations (declared, met, unmet, held, unmeasured)
  • obligation_states
  • held and unmeasured (id → reason), as their own buckets
  • uncited_cases: recorded translation cases no obligation cites. These aren't an error, because converter PRs add cases before their obligations exist, and they can't make a form match.
• The summary prints an obligations line. The baseline records each obligation's state and the retired ids. --against flags met → anything else as a regression, and so is dropping any obligation that retired-obligations.toml doesn't list. A missing obligations/ directory is a load error.
• Negative tests run through the loader, the measure, the CLI JSON and the diff. Among them: held counted as met, a form with a held obligation counted as matching, required = false without a reason, and an obligation half-met by two cases. That last test fails if "every cited case" is weakened to "any".

Figure (D272b: accepted as honest)

Measured on the landing head: main 9ecf9fe merged in (4c021f8), then the baseline re-recorded (bfecd63). Obligations: 3,407 declared, 1,034 met, 2,354 unmet, 19 held, 0 unmeasured, 0 load errors, 0 uncited cases.

• before (main 9ecf9fe baseline): 87 of 1,187 translatable forms matching
• after: 57 matching. 30 forms move to unmet_obligations; none are left at no_obligations.
• statements: 0 outcome changes (11,870 recorded).

The --against exception list (D272c: auditor-confirmed 30, record 2bd0e02)

translation-results --against main 9ecf9fe reports exactly these 30 form regressions, all Matching → UnmetObligations, and nothing else regresses. The other 3,407 changes are the obligation states the baseline didn't hold before (None → Met/Unmet/Held), plus 24 retired ids. The re-recorded baseline then checks clean: 0 regressions, 0 other changes.

Cause of the unmet case	Forms
Refused: CAST('…' AS BINARY), "a cast target type has no name" (12)	sf.functions.concat-ws, concat.concat, concat.concat-subscript-dot, left, right, length.len, length.length, lpad, rpad, position.position-in, substr.substr, substr.substring
Refused: the same BINARY cast and declared divergence json_fidelity at PARSE_JSON (7)	sf.functions.is-array, is-binary, is-boolean, is-integer, is-object, is-time, is-timestamp.is-timestamp-tz
Refused: unknown Snowflake type DECFLOAT (2)	sf.functions.ln, sqrt
Refused: TIMESTAMP_LTZ(9) → DATE has no proven conversion (2)	sf.functions.current-date.current-date, current-date.current-date.syntax
Declared divergence json_fidelity at ARRAY_CONSTRUCT (4)	sf.functions.array-compact, array-flatten, array-max, array-min
Obligations C landed after the first measure; at least one is unmet (3)	sf.functions.nullif (t003 held: it differs from Snowflake, fix on the post-launch list), sf.expressions.data-types-datetime.time, sf.expressions.data-types-logical

What would bring them back, ranked (candidate items)

1. A receipt-backed VARCHAR → BINARY cast rule (hex decoding), now in BINARY: the type, the VARCHAR→BINARY cast and byte-wise text functions (D284) #129 (approved, queued). This unblocks 19 cases: 12 forms come back outright, and the 7 IS_* forms then need only item 2.
2. json_fidelity at PARSE_JSON / ARRAY_CONSTRUCT for the shapes these cases use: 11 forms (7 also need item 1).
3. DECFLOAT as an input type: 2 forms.
4. TIMESTAMP_LTZ → DATE under a fixed session time zone: 2 forms.

Landing notes

• C's fix/translation-absent-forms is on main (df143dd), so the loader reports 0 load errors.
• Gate (D94) on 4c021f8, the exact merge with main 9ecf9fe:
  • make check-merge: exit 0, with the static checks and all 23 test binaries passing.
  • sqlquack verify --require-verified --strict-errors: 11,870 cases, 7,829 passed, 1,032 diverged as expected, 2,972 refused, 37 typed, 0 pending, 0 unproven, 0 failed; 220 of 220 equivalences and rules verified.
  • translation-results --against main: the 30 forms above and nothing else.
• bfecd63 only re-records coverage/translation/results-baseline.json. No test reads it, and results --against it gives 0 regressions and 0 other changes.
• Review note (auditor, round 2): retiring an unmet obligation can turn its form matching, so a retirement in retired-obligations.toml needs the same review as deleting the obligation.
• GitHub Actions is blocked by billing, so the local D94 gate counts.