Guided structured-result employee recipes for deliverable acceptance #189

Problem

Opt-in deliverable acceptance now checks pinned fields and typed claims, but a normal
prose-oriented employee may correctly describe a calculation without returning the
required structured field. The result correctly remains NEEDS_REVIEW; users still
need to understand and configure strict JSON output in a published Agent Version.

Private live qualification verified this difference: the prose profile stayed blocked,
while a separate explicit JSON profile returned the actual rate property and passed.
This is usability/output-contract work, not a request to loosen acceptance or extract
guessed quantities from arbitrary prose.

Proposed scope

• Provide an opt-in guided structured-result recipe when creating/publishing employees.
• Make required field/typed quantity examples visible at Task setup, with a clear warning
  when a selected employee's declared output contract cannot satisfy configured checks.
• Where a provider supports schema/JSON response formats, expose them through validated
  provider configuration without assuming cross-provider parity.
• Keep output/evidence pinned, legacy prose profiles unchanged, and unsupported/missing
  output blocked. A model's verdict must never grant acceptance.
• Add one repeatable Console walkthrough using a published JSON recipe; test missing
  fields, unsupported providers, unit mismatch and guided publication/version review.

Acceptance criteria

A new user can configure an employee and a numeric checked collaborative deliverable
from the Console without hand-writing a large system prompt. The actual structured
result passes only if the existing deterministic checks pass. No automatic natural-
language parsing fallback, change to default Tasks, or semantic-correctness claim.

Related: #188.