executor: resolve InputRequiredResult elicitations from static or REST answer sources

[rickcrawford]

What

Part 3 (final) of the 2026-07-28 runtime wiring. A server can answer a tools/call with an InputRequiredResult (SEP-2322); the recognizer, retry builder, and static fixture existed but nothing outside tests constructed them. This wires the live runner to satisfy an elicitation and retry, with a configurable, layered answer source (static map, REST endpoint), per the approved design.

Changes

• Suite config: a tool test declares input_responses (static id: value map) or input_responder (rest: { url, headers?, timeout_ms? }); a suite-level input_responder is the default. Test-level keys are mutually exclusive. Parsed config (InputResponderSpec/RestInputResponderSpec/InputResponseSpec) is kept separate from the runtime responder. Both root + embedded schemas (+ schema_validation) gain the keys and $defs.
• Runtime responder: a Responder (Static | Rest) built at plan time. Static delegates to the existing collect_responses; Rest POSTs the elicitation (full inputRequests, incl descriptions) to a proxy-aware reqwest client and validates the reply (non-2xx, invalid JSON, missing/duplicate/unknown ids, missing required, and kind). Uniform string/number/boolean kind validation for both sources.
• Live retry loop (executor/dispatch.rs::call_server): after a tools/call (and any task-handle poll), an InputRequiredResult with a configured responder is resolved, the retry is built off the transformed params, any task handle is re-polled, and it loops to the round cap. The final request id is returned for response-header assertions. Metamorphic follow-ups do not elicit.
• Plan build: resolves the effective responder (test override > suite default) with the run-wide proxy; a bad REST url/proxy fails plan building, not the first round. A REST responder marks the test cache-exempt (live IO).
• Observability: every round emits a redacted mcptest_core::elicitation trace event (round_started/answer_resolved/retry_dispatched/completed/failed) with ids, counts, responder kind, request id, and a stable requestState hash — never the raw state, answer values, or headers.
• Docs + example: elicitation.md, yaml-reference.md, reference/config.md; examples/async-tasks.yml book_flight now resolves from static input_responses to the final result (its own server, WOR-1473 guard).

Testing

• 5 parser tests, 12 elicitation tests (incl static + REST wiremock: happy path, HTTP error, unknown id, wrong-kind), 99 schema_validation tests.
• Offline end-to-end via examples/async-tasks.yml (examples gate); 30/30 under induced cargo-build load.
• fmt, clippy -D warnings, cargo doc -D warnings, em-dash, module-size green locally; whole workspace compiles all-targets.

Cassette-recorded elicitation answers remain a follow-up (WOR-943). This completes WOR-1383 (Parts 1 #72, 2 #73, 3 here).