feat(opencode): dispatch recipes + model flag

[slowdini]

Part of the OpenCode enhancement-parity track (see #113 for the baseline + progressive-enhancement model; per-enhancement wiring guide in docs/progressive-enhancements.md).

harnesses/opencode.toml currently declares no [dispatch] block and no [model] table, so run --harness opencode prints generic manual-opencode run guidance and --agent-model/--judge-model are recorded as provenance only (README support table footnote ¹).

Work

Add a [dispatch] block (exec_template, parallel_command_template, judge_command_template, next_steps_template, manifest_template, capture_prefix = "opencode") plus [model] flag = "-m" to harnesses/opencode.toml. These must land together: descriptor validation ties judge_command_template's $model_arg to a declared model flag.

Verified CLI facts (opencode v1.18.3 docs + packages/opencode/src/cli/cmd/run.ts)

• opencode run --dir <dir> resolves the project instance against <dir> and chdirs there, so staged .opencode/skills in the env are discovered.
• Piped stdin is appended to the message (Bun.stdin.text() when stdin is not a TTY) — recipes must detach with </dev/null>, same as codex.
• Headless permission asks are auto-rejected unless --auto is passed. The dispatch prompt lives outside the task env dir, and external_directory defaults to ask, so recipes need --auto (explicit deny rules are still enforced under --auto — relevant once the write guard lands).
• --format json streams raw JSON events (tool_use / text / step_finish / error) to stdout; there is no --output-last-message — the final message is recovered from text events by transcript ingest.
• -m takes models in provider/model format; the value passes through verbatim.

Proposed template shape

opencode run --dir <eval-root> --format json --auto{model_arg} \
  "Read the file at <dispatch_prompt_path> and follow its instructions exactly. When you finish, make your final response your closing summary." \
  </dev/null \
  > <outputs_dir>/opencode-events.jsonl \
  2> <outputs_dir>/opencode-stderr.log

(parallel + judge variants per the codex descriptor's structure; next_steps_template points at eval-magic ingest --harness opencode.)

Acceptance

• [dispatch] + [model] declared; eval-magic harness lint opencode clean
• Golden fixtures blessed (GOLDEN_BLESS=1 cargo test golden_) covering RUNBOOK/dispatch-manifest/next-steps for --harness opencode
• One live opencode run per the recipe in a scratch env verifies: staged skills discovered, events file ingests cleanly, model flag honored
• docs/opencode-notes.md updated; README support table flips Dispatch recipes and Model flag to ✅ and footnote ¹ is removed
• cargo test, cargo clippy --all-targets -- -D warnings, cargo fmt --check green

Sequenced after the transcript-ingest work so next_steps_template can reference ingest (avoids blessing the handoff text twice).