feat(cli): synthbench submit scaffold (refs #256)

[openclaw-dv]

Closes the scaffolding ask in #256.

Summary

• Adds synthbench submit-adapter — the vendor self-submission CLI requested for the SynthPanel-competitor onboarding path.
• Adds synthbench.adapter.Adapter abstract base + RandomAdapter reference impl for wiring-up tests.
• Adds docs/submit.md walking vendors through the three-step flow (adapter → submit → PR).

This is scaffold-only by design (per the issue scope). Inputs are validated end-to-end, but the produced submission.md + run.json are placeholders. The full eval pipeline, run-hash addressing, and leaderboard-PR auto-generation are deferred to follow-up issues against #256.

What's complete

• synthbench.adapter.Adapter abstract base — name, version, async respond(question, persona, context=None).
• synthbench.adapter.RandomAdapter reference impl (yes/no/Likert), seeded for deterministic tests.
• synthbench submit-adapter Click command with --adapter, --vendor, --vendor-version, --api-env-var, --suite (default core), --output-dir (default ./synthbench-submission/).
• Validation: adapter importable (exit 2), Adapter subclass present (exit 2), API env var set without reading it (exit 3), no-arg constructible (exit 2).
• Placeholder artifact writer (submission.md + run.json) with the eventual leaderboard PR URL + stubbed verification curl.
• Filesystem-path and dotted-import adapter resolution (vendors will use file paths; in-tree adapters use dotted).
• One-line wire-up in src/synthbench/cli.py — no other existing files touched.
• 5 tests in tests/test_cli_submit_adapter.py covering --help, missing module, missing env var, happy path, and filesystem-path import. All pass (0.03s). Existing test_cli_submit.py + test_cli_run_submit.py (29 tests) still pass.
• docs/submit.md with worked RandomAdapter example + exit-code table.

TODO (tracked as follow-ups to #256)

• Drive Adapter.respond through the core suite runner (the actual eval). Stubbed today.
• Content-addressed run-hash over (adapter id, suite manifest, persona seeds, raw responses). run_hash: null in the scaffold.
• Auto-generate the leaderboard PR body and open it via gh. Scaffold just prints the target URL.
• CI workflow validating submission PRs (run-hash reproducibility + registry of accepted adapter shapes).

Architectural decisions punted to a human eye

1. Command name collision with existing synthbench submit. The repo already has a submit subcommand that posts a pre-computed result.json to the hosted SynthBench API (tests/test_cli_submit.py pins that contract). The new vendor-adapter flow is a different verb (vendor brings code, harness drives the eval), so I registered it as submit-adapter to honor the "don't restructure outside the new files" constraint. Options for final naming:

   • Keep submit-adapter permanently.
   • Rename the existing submit → submit-result (or push) and reclaim submit for the adapter flow (breaking change for current API users — needs deprecation cycle).
   • Make submit a subgroup with submit result and submit adapter subcommands (also breaking).
     I have no opinion on which; flagging for Wesley/maintainer call.

2. Module location. The directive specified src/synthbench/cli/submit.py, but cli is a single 2577-line module here, not a package. Converting it to a package would be a much bigger refactor (touches every from synthbench.cli import main import in the test suite). I put the new logic in src/synthbench/submit_adapter.py instead and wired it via one import in cli.py. If maintainers want the CLI broken up into a package eventually, this is a clean seam.

3. Tests live at tests/test_cli_submit_adapter.py, not tests/cli/test_submit.py. The existing test layout is flat (tests/test_cli_*.py); creating a tests/cli/ subdir would be inconsistent.

4. RandomAdapter ships in synthbench.adapter rather than under tests/. Reasoning: vendors are explicitly going to point --adapter synthbench.adapter at it as the known-good smoke target while debugging their setup. Test-only placement would make that flow fragile. It's clearly marked "not a baseline" in the docstring.

Test plan

• pytest tests/test_cli_submit_adapter.py — 5/5 pass
• pytest tests/test_cli_submit.py tests/test_cli_run_submit.py — 29/29 pass (no regressions to existing submit flow)
• synthbench submit-adapter --help exits 0 with the documented options
• synthbench --help shows both submit and submit-adapter
• ruff check clean on all new files

Refs #256