Schema-driven generation and typed-outputs hardening (follow-on to #272)

[anticomputer]

Follow-on work after the typed-outputs / grammar-maturity PR (#272). Now that the outputs contract is standard JSON Schema (Draft 2020-12) and capture selects tool-result vs response text, the schema-driven pieces below are the natural next steps. Split out of #272 to keep that PR reviewable as a coherent unit.

Schema-driven generation (highest value)

• Pass the outputs JSON Schema to each backend's native structured-output surface (OpenAI response_format, Anthropic tool input_schema; copilot_sdk falls back to validate-only) so the model is constrained to emit conforming output instead of relying on prompt prose. This is what makes capture: response + outputs reliable rather than best-effort.

Retry-to-repair

• On strict-validation failure, re-prompt the model with the validation error (bounded) so strict validation is robust rather than brittle. Knob: outputs_retries.

Reusable schemas

• A shared-schema mechanism (top-level schemas: / document type) so a contract can be defined once and referenced across tasks.

format policy

• Decide whether to enforce JSON Schema format (jsonschema does not by default) and enable a FormatChecker if so.

Multi-model fan-in observability

• Surface per-branch validation errors (not just result: null) and add a consensus/merge helper over per-model typed results.

Friendlier validation errors

• Wrap jsonschema pointer-path messages into field-oriented messages for YAML authors.

CLI mode-flag ergonomics

• Reject ambiguous combinations of the mutually-exclusive terminal modes (--schema / --manifest / --lint / --list-models) with a clear error, instead of silently applying precedence, while still allowing each mode's legitimate modifiers (e.g. --lint with -t/-m).

Deferred (not planned)

An opt-in coercing mode: strict validation is intentional; the preferred path is generation + repair above.