feat(opencode): sub-agent dispatch parity via agents and the task tool

[thewrz]

This was written agentically; verify its assertions:

Why

parallel-issues and review-remote-pr dispatch implementation workers through a harness-specific spawn path (Codex spawn_agent, Claude subagents). OpenCode's verified dispatch surface is different: markdown-defined agents (mode: subagent) invoked via the task tool, with per-agent model: provider/model-id and permission scoping. Without a mapped path, an OpenCode session can only take the worker=self (spawn unavailable) degradation for every issue.

What

• Ship an agentkit-worker agent definition (.opencode/agents/, mode: subagent) whose model resolves from the repo's .agent/config.env worker declarations through the S2 spawn-contract mapping; document how the composed issue-lead prompt (from compose-worker-prompt.sh) is delivered through a task-tool invocation.
• Extend spawn-contract.md's dispatch section with the OpenCode path: task tool = spawn; context isolation semantics (verify at implementation whether a task-tool subagent inherits conversation context — the fork_context:false equivalent must be proven, not assumed); nesting behavior (can a subagent invoke the task tool? verify and pin, mirroring the existing "nesting is blocked" finding on Codex).
• Concurrency: verify whether OpenCode advertises a session concurrency cap; if none is discoverable, the skills' cap probe must return the documented no-cap path rather than stalling.
• Completion-report parity: worker returns branch/SHA/diffstat/log-path exactly as other harnesses; worktree-commit.sh (post-fix(scripts): worktree-commit.sh owns commit attribution instead of trusting callers #313) derives the trailer when the worker omits it.

Acceptance

• One end-to-end dispatched worker on OpenCode: composed prompt in, pushed branch + completion report out (transcript evidence in PR)
• Context-isolation and nesting behavior verified and pinned in spawn-contract.md with evidence, not assumption
• Cap probe behavior on OpenCode documented and tested
• Completion table records the actual provider/model-id that ran
• Unit tests pass; CI green

Blocked by the harness-identity (S2) and skills/commands (S4) slices.

🤖 Co-authored by Claude Fable 5.