fix(native-eval): bootstrap only the assigned harness

[steipete]

Problem

Fleet knows the harness assigned to each lease, but hydration discarded that selection. The bootstrap script consequently installed OpenClaw, Codex, Claude Code, and Hermes on every machine, even for a single-harness campaign.

Change

Pass the run's harness into bootstrap and require exactly one supported selection. Keep Docker, Node, uv, and LiteLLM shared; install and probe only the selected harness. Move uv's managed-Python/cache configuration out of Hermes setup so the shared LiteLLM environment remains self-contained when Hermes is not installed.

The manifest retains the selected harness's existing version key and adds harness; uninstalled harness keys are absent. Manual bootstrap now requires a harness argument. The research runbook and changelog document this behavior.

Fixes #65.

Proof

• Before the fix, all 11 new regression cases failed on both macOS and Blacksmith Testbox. They execute the shell script with harmless package-manager recorders and exercise fleet propagation for all four harnesses.
• After the fix, 50 focused tests pass locally. Full Blacksmith suite: 455 passed, 5 skipped, with one existing Gradio warning; full Ruff and wheel build pass.
• Python 3.11 and 3.12 CI pass, including runtime-contract tests and wheel-data validation.
• Fresh AWS Linux bootstrap with bootstrap_beast.sh openclaw: exit 0 in 171.542 seconds. Installed OpenClaw 2026.7.1-2 (0790d9f), Node v22.23.1, and LiteLLM 1.93.0. Ran the real OpenClaw CLI and LiteLLM Python environment, asserted Codex/Claude Code/Hermes directories absent, and verified the manifest contains only OpenClaw plus shared versions.
• Installed ClawBench CLI smoke, shell syntax, applicable pre-commit hooks, and isolated Codex autoreview pass. Autoreview used its default P0 reporting threshold.

The AWS smoke followed Testbox teardown/concurrency trouble and a Testbox console-script PATH failure after its suite and wheel build had passed. This is installer proof, not a model-evaluation run or a measured bootstrap speedup. Actual installation was exercised for OpenClaw; the other three selections have executable shell regression coverage.

[clawsweeper]

🦞👀
ClawSweeper picked this up.

Pull request received. I will update this pull request when review starts.

[clawsweeper]

ClawSweeper status: review started.

I am starting a fresh review of this pull request: fix(native-eval): bootstrap only the assigned harness This is item 1/1 in the current shard. Shard 0/1.

This placeholder means the worker is alive and reading the current context. I will edit this same comment with the actual review when the claws are done clicking.

Crustacean status: shell secured, claws on keyboard, evidence pebbles being sorted.