feat(opencode): dispatch recipes + model flag

[slowdini]

User-facing change

run --harness opencode now generates copy-pasteable dispatch recipes everywhere the codex/claude harnesses already did: RUNBOOK.md, dispatch-manifest.md, the post-run next steps, and the post-ingest judge handoff all carry opencode run --dir <eval-root> --format json --auto commands (including the shared jq/xargs parallel and judge variants). --agent-model / --judge-model now actually select models via -m instead of being recorded as provenance only.

Before:

Next: iterate the tasks[] array in dispatch.json and dispatch each task with `opencode run --format json`, saving stdout as the task's outputs/opencode-events.jsonl (copy-pasteable recipes are not yet wired). Model selection was recorded as provenance, but the OpenCode adapter has no CLI model flag wired yet. Then run `ingest …`.

After:

Next: iterate the tasks[] array in dispatch.json and dispatch each task with:
opencode run --dir <eval-root> --format json --auto -m opencode/claude-haiku-4-5 \
  "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
Then run `ingest …`.

What changed

• harnesses/opencode.toml: full [dispatch] block (exec_template, parallel_command_template, judge_command_template, next_steps_template, manifest_template, capture_prefix = "opencode") plus [model] flag = "-m" — landed together because descriptor validation ties the judge template's $model_arg to a declared model flag. No schema changes (all fields pre-existing). No {guard_args} references — the guard stays with feat(opencode): write guard via project plugin (opencode-plugin engine) #155.
• Golden fixtures: re-blessed runbook/manifest/next-steps; new opencode/judge-recipe.golden.md (opencode added to the judge golden sweep).
• Tests: flipped the has_dispatch_recipes pin; the dispatchless and modelflag-less preflight-warning pins moved from opencode to a BYOH user descriptor (every built-in now wires both); new unit pin for the exec recipe's --dir/--auto/-m splicing; new run_writes_headless_runbook_for_opencode integration test; deleted opencode_run_warns_missing_dispatch_recipe.
• Docs: README support table flips Dispatch recipes and Model flag to ✅ and footnote ¹ is removed; docs/opencode-notes.md records the verified CLI facts (v1.18.3) including the --auto/explicit-deny interaction; src/cli/help.rs drops the stale "assemble records manually" OpenCode example (leftover from before feat(opencode): transcript ingest via the opencode-events parser #156).

Verification

• cargo test — 479 lib + 99 cli + 81 run, 0 failures
• cargo clippy --all-targets -- -D warnings — clean
• cargo fmt --check — clean
• eval-magic harness lint opencode — all checks passed
• Live (opencode v1.18.3, scratch env): one opencode run per the recipe exited 0; the events file shows the staged skill discovered and invoked under its staged slug via the native skill tool; eval-magic ingest assembled run.json (3 invocations) + timing.json (tokens/duration from the transcript) and the __skill_invoked meta-check passed via transcript_check; -m definitely/bogus-model exits 1 (flag plumbed). An operator-level edit: deny rule stayed enforced under --auto and blocked final-message.md — record-runs fell back to the transcript's final text, as designed (noted in opencode-notes.md).

Closes #153