fix(cli): wire --terminal-reviewer-provider/acp-command through to config

[stephanbrez]

Problem

The --terminal-reviewer-provider and --terminal-reviewer-acp-command flags were accepted by click but silently dropped in _resolve_selection: ProviderSelection had no terminal_reviewer fields, so the values never reached env() and were never written to .mcp.json or .codex/config.toml.

HarnessConfig already reads ZENITH_TERMINAL_REVIEWER_PROVIDER and ZENITH_TERMINAL_REVIEWER_ACP_COMMAND with a full terminal_reviewer -> validator -> worker cascade — only the write side was broken.

Fix

• Add terminal_reviewer and terminal_reviewer_acp_command fields to ProviderSelection, with resolved_terminal_reviewer / resolved_terminal_reviewer_acp_command properties mirroring the existing validation_worker pattern.
• Update env() to emit the terminal-reviewer vars only when they differ from the validator's (the cascade parent), matching how the validator compares against the worker.
• Include the terminal reviewer in providers() and the skill dirs so its assets get installed when it's a distinct provider.

Tests

• terminal-reviewer provider + acp-command land in .mcp.json env.
• unset flags emit no terminal-reviewer env keys (cascade handles it).
• three distinct providers (claude/codex/hermes) all write their env vars in a single init.
• round-trip: a distinct terminal reviewer written by env() is read back intact by HarnessConfig.discover().
• round-trip: with no explicit terminal reviewer, env() omits the vars and discover() reconstructs the validator as the cascade fallback.

All three CI checks pass locally: ruff check ., mypy src, pytest -q (207 passed, 7 skipped).

Note: issue creation is restricted on this repo, so per CONTRIBUTING.md's fallback for small fixes this PR is filed directly.

🤖 Generated with Claude Code