feat(onboard): chat-based conversational setup assistant as default zeroclaw onboard

[JordanTheJet]

Summary

• Base branch: master
• What changed and why:
  • Revives zeroclaw onboard (previously a deprecation stub) as a conversational chat-based setup assistant, ported from OpenClaw's onboard --modern ("Crestodian"). It gives new users a plain-language path to inspect state and stand up their first agent without memorizing flags or section names.
  • Adds src/onboard/ with a clean split: overview (resolved-on-demand state snapshot), operation (closed command vocabulary + deterministic parser), planner (LLM fallback), execute (approval-gated executor), mod (interactive loop + --message/--json modes).
  • Deterministic-first by design: input is matched against a closed vocabulary before any LLM is consulted; the planner is reached only on unrecognized input and its output is re-parsed through the same parser, so model text can never execute outside the vocabulary. With no model configured (first run), it runs deterministic-only.
  • All user-facing output routes through Fluent (cli-onboard-* keys in locales/en/cli.ftl) per the cli_fluent_coverage architecture gate.
• Scope boundary: Does not change quickstart (reused as the setup path), the daemon/gateway/agent runtimes, the config schema (no new fields), or provider behavior. The Tauri/web/zerocode onboarding surfaces are untouched.
• Blast radius: Limited to the zeroclaw onboard CLI command. All config writes flow through the existing sanctioned set_prop_persistent + save_dirty path or the quickstart apply path; reads reuse doctor::run / run_configured_models and the provider factory. No new shared/cached state.
• Linked issue(s): None.
• Labels: Suggested: type: feature, risk: low, size: M.

Validation Evidence (required)

cargo fmt --all -- --check                  # clean
cargo clippy --all-targets -- -D warnings   # clean
cargo test                                  # all suites pass

• Commands run and tail output:

cargo clippy --all-targets -- -D warnings: Finished ... in 30.07s (no warnings).

cargo test (full root battery, all green):

test result: ok. 230 passed; 0 failed   (architecture — includes cli_fluent_coverage)
test result: ok. 281 passed; 0 failed   (bin: onboard parser + CLI)
test result: ok. 188 passed; 0 failed
test result: ok. 159 passed; 0 failed
test result: ok.  19 passed; 0 failed
test result: ok.   9 passed; 0 failed
test result: ok.   3 passed; 0 failed

• Beyond CI — what did you manually verify? Built the debug binary and exercised it against a live config: onboard --json and onboard --message "status" resolve agents/providers (with per-provider key presence)/gateway reachability; persistent ops (set up an agent) print an approval plan and apply nothing without approval. Live planner verified with a real model: a free-form request ("give me the lay of the land") routed through the LLM planner to overview (executed), and "what should I do first?" routed to setup (plan printed, correctly gated). Localized output (help, agents, gateway status, plans) renders with all Fluent keys resolved.
• If any command was intentionally skipped, why: None — the full cargo test battery was run.

Security & Privacy Impact (required)

• New permissions, capabilities, or file system access scope? No — reads the existing config path and writes config only through the same sanctioned setter zeroclaw config set uses.
• New external network calls? Yes — a bounded (~2s) health probe to the user's locally-configured gateway /health, and the LLM planner call to the user's already-configured model provider (only on free-form input when a keyed provider exists). Both target endpoints the user already configured; no new third-party endpoints.
• Secrets / tokens / credentials handling changed? No — secrets are never printed; the approval-plan redaction uses the schema's canonical Config::prop_is_secret; keys are read from existing config and handed to the existing provider factory.
• PII, real identities, or personal data in diff, tests, fixtures, or docs? No.
• If any Yes, describe the risk and mitigation: Planner output is untrusted model text — re-validated through the deterministic parser (cannot execute outside the closed vocabulary) and stripped of control/escape characters before display; all state-changing operations require explicit approval.

Compatibility (required)

• Backward compatible? Partially — zeroclaw onboard changes behavior: it was a deprecation stub and now launches the chat assistant. quickstart is unchanged.
• Config / env / CLI surface changed? Yes (CLI only) — onboard now accepts --message/--json/--yes and no longer accepts the removed legacy flags (--quick, --force, --reinit, --api-key, --model-provider, --model, --memory, --<section>-only, positional section subcommands). No config or env changes.
• If No or Yes to either: exact upgrade steps: Scripts that invoked zeroclaw onboard with legacy flags should switch to zeroclaw quickstart (guided agent creation) or zeroclaw config set <path> <value> (headless). Bare zeroclaw onboard now opens the interactive assistant (TTY required); use --message "<request>" or --json for non-interactive use.

Rollback (required for risk: medium and risk: high)

Low-risk: git revert <sha> restores the prior deprecation stub. The change is isolated to the onboard command path.

Supersede Attribution (required only when Supersedes # is used)

N/A — not superseding any PR.