docs(spec): state the residual bound the changeset states, and point the ADR-0087 marker at the right bullet

F1 — `packages/spec/src/data/numeric-column-representation.ts`'s "residual
bound" section said only that magnitudes below 1e-30 round to zero, while the
shipped changeset states the column keeps 30 fractional digits, that precision
loss BEGINS around |x| < 1e-13 and is TOTAL below 1e-30, and that magnitudes at
or above 1e35 are refused. That docblock is the canonical home both
`sql-driver.ts` and `os generate migration` point readers at ("the spec module
carries the measurements, the ruling and the residual bound"), and it says of
itself that the bound is "stated here so no reader has to rediscover it" — so
it now states both ends of the bound in the changeset's own terms.

F2 — `.changeset/numeric-column-representation.md` said "Three consequences"
above four bullets, and its ADR-0087 marker claimed "bullet 3 IS a
prescription" where bullet 3 is the read-seam bullet and the prescription is
bullet 4 (the `NOT NULL` / `storage.notNull` one). The marker is the auditable
artifact the gate prints verbatim, so a wrong internal pointer defeats its
purpose. Exactly one `adr-0087:` marker, same category and same argument.

Docblock and changeset prose only: no behaviour change, no schema move, no
test edit.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017Js5kTpTtxieBjPyScgxJ3

Author: claude

.changeset/numeric-column-representation.md

@@ -36,7 +36,7 @@ the write-time contract the record validator enforces, and binding the DDL to it
 post-deploy tightening a destructive migration.
 
 **BREAKING** — new columns only; no existing column is retyped, no migration is planned, and no
-backfill runs. Three consequences to know before creating new tables:
+backfill runs. Four consequences to know before creating new tables:
 
 - `rating` is an INTEGER column, and the two server dialects dispose of a fractional star count
   DIFFERENTLY — do not read one answer for both. PostgreSQL REFUSES `4.5` outright, where a
@@ -67,4 +67,4 @@ backfill runs. Three consequences to know before creating new tables:
 SQLite emits byte-identical DDL for the six exact-decimal members: knex compiles both
 `table.decimal(name, p, s)` and `table.float(name)` to the same `float` column there.
 
-<!-- adr-0087: not-required (no-migration-prescription) Claimed on a POSITIVE argument, not on the detector finding nothing — the failure mode this gate's own docblock names (#8277). Stated plainly: bullet 3 IS a prescription, and it is a prescription for a SOURCE AUTHOR, not for a metadata upgrader, which is the distinction ADR-0087's D8 addendum says this category cannot mechanically tell apart. The ledger serves `objectstack migrate meta`; the only ledger entry this change could carry is the `field-required-notnull-explicit` conversion, and that conversion was WITHDRAWN by maintainer ruling on 2026-09-08 (decision batch #85, #16693/#16890) on the ground that stamping `storage.notNull` wherever `required: true` appears is the implication ADR-0113 abolished — `packages/spec/src/conversions/registry.ts` now carries a tombstone saying re-adding one is the mistake it exists to stop. So `registered` is FORBIDDEN here, not merely unnecessary. The other four are closed on facts: the bumped packages publish (not `unpublished`); no id pre-dates the base (not `already-registered`); no named symbol is a non-metadata runtime interface (not `runtime-interface-only`); and `type-surface-only` fails its predicate 2, since this diff adds a module under `packages/spec/**`. The numeric half prescribes nothing at all — no spec key, no export and no config field is removed or renamed, existing sources parse and publish unchanged, and existing columns are untouched by the ruling that authorized this (「不考虑现有数据」). ⚠️ The residual is declared rather than hidden: the vocabulary has no category for a source-author prescription the ledger must not carry, which is D8's blind spot reached from a second direction; raised for the maintainer in the PR report rather than resolved by dropping the BREAKING banner. -->
+<!-- adr-0087: not-required (no-migration-prescription) Claimed on a POSITIVE argument, not on the detector finding nothing — the failure mode this gate's own docblock names (#8277). Stated plainly: bullet 4 (the `NOT NULL` one) IS a prescription, and it is a prescription for a SOURCE AUTHOR, not for a metadata upgrader, which is the distinction ADR-0087's D8 addendum says this category cannot mechanically tell apart. The ledger serves `objectstack migrate meta`; the only ledger entry this change could carry is the `field-required-notnull-explicit` conversion, and that conversion was WITHDRAWN by maintainer ruling on 2026-09-08 (decision batch #85, #16693/#16890) on the ground that stamping `storage.notNull` wherever `required: true` appears is the implication ADR-0113 abolished — `packages/spec/src/conversions/registry.ts` now carries a tombstone saying re-adding one is the mistake it exists to stop. So `registered` is FORBIDDEN here, not merely unnecessary. The other four are closed on facts: the bumped packages publish (not `unpublished`); no id pre-dates the base (not `already-registered`); no named symbol is a non-metadata runtime interface (not `runtime-interface-only`); and `type-surface-only` fails its predicate 2, since this diff adds a module under `packages/spec/**`. The numeric half prescribes nothing at all — no spec key, no export and no config field is removed or renamed, existing sources parse and publish unchanged, and existing columns are untouched by the ruling that authorized this (「不考虑现有数据」). ⚠️ The residual is declared rather than hidden: the vocabulary has no category for a source-author prescription the ledger must not carry, which is D8's blind spot reached from a second direction; raised for the maintainer in the PR report rather than resolved by dropping the BREAKING banner. -->

packages/spec/src/data/numeric-column-representation.ts

@@ -98,12 +98,19 @@
  * ## The residual bound, stated rather than assumed
  *
  * An exact-decimal column is bounded where a float is not, so this table is
- * not lossless in every direction. Magnitudes below 1e-30 round to zero and
- * magnitudes at or above 1e35 are REFUSED, where today's `real` keeps about
- * seven significant digits out to ~1e38. Two things make that the right
- * trade: a refusal is loud and a silent rounding is not, and the sql format's
- * `numeric(18,2)` already refuses everything at or above 1e16 today. It is a
- * bound, and it is stated here so no reader has to rediscover it.
+ * not lossless in every direction, and the bound has two ends. Downward, the
+ * column keeps 30 fractional digits, so a magnitude whose significant digits
+ * run past the 30th decimal place loses the tail silently:
+ * `1.2345678901234567e-15` stores as `0.000000000000001234567890123457`.
+ * Precision loss therefore BEGINS around |x| < 1e-13 — where a double's ~17
+ * significant digits first reach past the 30th decimal place — and is TOTAL
+ * below 1e-30, where nothing is left and the value rounds to zero. Upward,
+ * magnitudes at or above 1e35 are REFUSED (the 35 integer digits that
+ * 65 - 30 leaves), where today's `real` keeps about seven significant digits
+ * out to ~1e38. Two things make that the right trade: a refusal is loud and a
+ * silent rounding is not, and the sql format's `numeric(18,2)` already
+ * refuses everything at or above 1e16 today. It is a bound, and it is stated
+ * here so no reader has to rediscover it.
  *
  * ## SQLite, per type — the constraint the report raised, answered
  *