fix(opencode-agents): skip judges and record structured history

[apucacao]

Judges still match this handler because provides_for stays ("*", "agent"). A judge evaluation now raises before events() runs, so the judge prompt does not open the OpenCode session and the judge does not score.

_input_messages records block-array history on the invoke_agent span, using the same text flattening as the other handlers, and does not append an empty user_input.

provides_for is unchanged. This does not change #159.

---

Note

Overview
The OpenCode handler now detects judge evaluations via OpenTelemetry context and raises before events() runs, so a judged config does not reopen the OpenCode session on the judge prompt or produce a bogus score. Callers can still use skip_judges=True on config() to avoid selecting a judge attempt entirely. Docs and examples no longer treat skip_judges=True as required for normal agent streaming.

_input_messages now builds invoke_agent span input with compose_history and content_to_text, matching other handlers: block-array turns are flattened to text, system messages are omitted from gen_ai.input.messages, and an empty user_input does not add a blank user turn.

Tests cover judged streaming (events called once, no judge_results) and block-array history on the span.

^{Reviewed by Cursor Bugbot for commit e54111d. Bugbot is set up for automated code reviews on this repo. Configure here.}