Adapter function metrics record success for contract-level IOContract mismatches #1611

Summary

mellea.adapter_function.invocations/.parse_failures mis-record one class of failure for adapter functions called through call_intrinsic: a contract-level IOContract mismatch on output that is valid JSON.

Background

PR #1609 (issue #1560) wired adapter_function_invocation_complete firing for the Embedded/Granite Switch binding. It fires the outcome once each backend's granite_formatters_processing closure resolves — success if the response parses as JSON, schema_error if it doesn't, error otherwise.

That closure runs before call_intrinsic (mellea/stdlib/components/intrinsic/_util.py:171) calls adapter.io_contract.parse(result_str) — the actual contract check (required fields, types, shape). So when a response is syntactically valid JSON but fails that contract check, AdapterSchemaMismatchError is raised to the caller — but the invocation was already recorded as outcome="success".

Impact

• The caller is unaffected: AdapterSchemaMismatchError propagates uncaught through every wrapper (rag.py, core.py, guardian.py) straight to application code. No silent failures.
• The metric is wrong: .parse_failures undercounts this specific failure mode, and .invocations{outcome="success"} overcounts it. A model that drifted into returning valid-but-wrong-shape JSON would look 100% healthy on a dashboard built from these metrics while every call actually fails downstream.
• Scope is narrow: only applies to output that's valid JSON but violates the adapter's declared schema — not malformed JSON (already schema_error) and not a generation/network failure (already error).
• Applies to local_file/PEFT too, not just embedded — AdapterMixin.adapter_scope's outcome closes at generation return, before parsing or contract validation ever run, so this binding never records anything for this failure mode either.

Proposed fix

Move (or add) the invocation-complete fire point into call_intrinsic, after io_contract.parse() resolves, so the recorded outcome reflects contract validation too. Complications:

• call_intrinsic is shared by both bindings — this needs handling that's correct for local_file as well as embedded, not an embedded-only patch.
• call_intrinsic is synchronous; hook dispatch (invoke_hook) is async. Needs the same sync→async bridging adapter.py's _fire_invocation_complete already does via _run_async_in_thread — the embedded fire helper introduced in fix(backends): fire adapter_function_invocation_complete for embedded adapter calls #1609 (_fire_embedded_invocation_complete) deliberately avoids that bridging since it's always called from an already-running coroutine, so this can't just reuse it as-is.
• Need to decide whether the backend-level success fire (in the closures) should be suppressed/deferred entirely in favour of this one, to keep the "fires exactly once per invocation" invariant intact.

cc @ajbozarth