docs(spec): correct the 17.4.0 duration-keys erratum (#18076)

The published packages/spec/CHANGELOG.md 17.4.0 entry, under "Keys
deliberately left alone", said the burn-rate `window` and the
OpenTelemetry exporter `timeout` name no unit anywhere in their prose
and are outside check:duration-unit-keys's population, with pin tests
holding the bare spellings. All three claims are false now: #17783
renamed the burn-rate window to `durationSeconds` and #17785 renamed
the exporter timeout to `timeoutMs` (with its `exportTimeout` sibling
to `exportTimeoutMs`), both tombstoned with `retiredKey()`, and the
pin tests that held the bare names were rewritten into rename-refusal
pins. This corrects the sentence to the post-#17783/#17785 truth and
appends one erratum line, matching the ef47459 (#17896) precedent's
shape.

Clause-②: no

skip-changeset: no package publishes anything from this change — the
edit is prose inside an already-compiled CHANGELOG.md section plus an
erratum line.

Author: claude

packages/spec/CHANGELOG.md

@@ -2862,11 +2862,25 @@
   milliseconds where its `resetTimeoutSeconds` sibling is seconds, so the two must
   not be migrated as if they were one unit. `MetricExportConfig.batch.size` is a
   record count and `QueueConfig.rateLimit.max` is a task count: neither is a
-  duration, so neither has a unit to carry. `ServiceLevelObjective.errorBudget`'s
-  burn-rate `window` and the OpenTelemetry exporter `timeout` name no unit
-  anywhere in their prose, so both are outside the gate's population entirely.
-  Pin tests assert each of these, so a later sweep cannot read this card as
-  "every duration-shaped number on these files".
+  duration, so neither has a unit to carry. As published, this entry continued:
+  "`ServiceLevelObjective.errorBudget`'s burn-rate `window` and the
+  OpenTelemetry exporter `timeout` name no unit anywhere in their prose, so
+  both are outside the gate's population entirely. Pin tests assert each of
+  these, so a later sweep cannot read this card as 'every duration-shaped
+  number on these files'." Both keys were renamed in the same major: #17783
+  renamed the burn-rate `window`
+  (`ServiceLevelObjective.errorBudget.burnRateWindows[].window`) to
+  `durationSeconds`, and #17785 renamed the OpenTelemetry exporter `timeout`
+  (`OpenTelemetryCompatibility.exporter.timeout`) to `timeoutMs`, with its
+  `batch.exportTimeout` sibling moving to `exportTimeoutMs` the same way.
+  Both are inside the gate's population now, each tombstoned with
+  `retiredKey()` naming the new spelling, and the pin tests that held the
+  bare names were rewritten by the same two PRs into rename-refusal pins.
+  
+  *Erratum, 2026-09-13 — #17783 and #17785 renamed the two keys this entry
+  named as gate-exempt; both are tombstoned and inside the gate's
+  population as of `@objectstack/spec` 17. (Corrected after publication,
+  #PRPLACEHOLDER18076.)*
 - fb77aa5: feat(spec)!: a `tree` field's `reference`, when present, must name the declaring object — any other target is refused at parse (#14892)
 
   <!-- adr-0087: not-required (no-migration-prescription) No authorable key is renamed, retired or re-typed: `reference` keeps its name, its type and its optionality on a `tree` field, and every self-referencing or reference-less `tree` parses byte-identically to before. What is newly refused is a `tree` whose `reference` names a different object — a shape no prose surface ever documented, that no runtime reader consumed as a cross-object link, and whose one in-repo author (the showcase field zoo) hedged in its own label. The remedy is authoring intent, not a mechanical rewrite: the author decides whether the field is this object's own hierarchy (a self-reference, or no `reference` at all) or a link to another object (a `lookup`), so `objectstack migrate meta` has nothing to rewrite and this changeset carries no rewrite instructions. -->