#4525 asked the spec to model `required` on `DecisionOutputDef` so the
contract declares what `ApprovalService.decide()` already enforces. The
modelling itself landed three days before the issue was filed — cd6b9f2
(`feat(approvals): decisionOutputs may be declared required`, objectui#2955)
added the Zod key, its TSDoc, the `strictUnknownKeyError` known-key entry,
the `normalizeDecisionOutputs` pass-through, the authorable-surface baseline
row and the generated docs table. The issue was recorded from objectui's
ledger burn-down, whose derived-type comment ("the spec does not model it
yet") was written against an older spec and was already stale.

What was genuinely missing is the pin. `DecisionOutputDefSchema` is
`.strict()`, so dropping the key would turn every author's `required: true`
into an unknown-key rejection — and the only thing standing in the way was
`authorable-surface.json`, a REGENERATED baseline that a `gen:schema` run
rewrites without comment. Verified by mutation: commenting the key out left
all 40 pre-existing cases green, including the `normalizeDecisionOutputs`
ones, because that normalizer is hand-written and never consults the schema.

Three cases close it — the key parses and round-trips, it stays optional
(absent must not become `required: false`; the parsed shape is what ships to
every decision UI on `decision_output_defs`), and a non-boolean is rejected
rather than coerced. That last one is the AI-authoring case: the runtime
compares `d.required === true`, so a coerced `'true'` would declare a
constraint the server then never enforces — the same declared-vs-enforced
gap #4525 exists to close, one layer down.

No behaviour change; no changeset (the feature shipped with its own in
cd6b9f2).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012C2cd7tL8QDoZ2QKN3djJ5

The "Check Changeset" gate counts changesets ADDED by the PR relative to
base, so pointing at cd6b9f2's `.changeset/required-decision-outputs.md`
(which shipped the feature itself) does not satisfy it. An empty-frontmatter
changeset is the sanctioned form for a PR that bumps no package: the diff is
one test file pinning a spec key that already exists, so there is nothing for
a consumer to read in a CHANGELOG.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012C2cd7tL8QDoZ2QKN3djJ5