feat: add labeled evaluation and observation hooks

[gudcks0305]

Add opt-in completion metadata on provider builders and a labeled evaluation example for choosing probability thresholds.

• .observer(...) reports one logical evaluation's duration, requested/returned model, question count, usage, outcome, and bounded error kind/status. Events exclude keys, headers, endpoints, state, questions, answers, raw responses, and exception messages. Observer failures do not alter inference or cancellation; transport retries remain one observation.
• LabeledEvaluationExample defaults to eight synthetic offline rows without credentials or network calls. Explicit --live mode sends only state, keeps gold labels local, and supports labeled JSONL files. Threshold reports distinguish errors and abstentions, compute coverage over all rows, and report accepted accuracy as N/A when nothing is accepted.
• Add usage/lifecycle/metric documentation and mark these features unreleased. No logging framework, Micrometer binder, provider module, or default model fallback is introduced.

Validation:

• ./mvnw -q verify: 127 tests passed, no failures/errors/skips.
• Offline example ran successfully; its figures are synthetic demonstrations, not measured model quality.
• Observation guide Java example compiled with javac --release 17; local documentation links checked.
• Independent read-only review found no confirmed defects.

The compatibility documentation and repository AGENTS.md were merged separately in #15. This PR does not claim successful live inference on additional gateways.