feat(run): auto-arm the write guard; scope preflight warnings

[slowdini]

Closes #126.

What changed

Issue #126 asked that every harness enhancement be detected and provided automatically, with warnings naming the limitation wherever a harness lacks one. The preflight already implemented the warn-and-degrade model; this PR closes the remaining gaps:

• The write guard now arms automatically on guard-capable built-in harnesses whenever staging is active. A new --no-guard flag opts out; --guard remains as the explicit form (it warns where auto-arm quietly stays off — guardless harness, --no-stage — and is still rejected on user-descriptor-only harnesses, since the guard fails open). This replaces the README's "arm --guard unless the user opts out" agent-etiquette with actual default behavior. RunOptions.guard is tri-state (None = auto) and the preflight resolves it; the run plan prints guard: armed … (--no-guard to opt out).
• The transcript warning is scoped to configs that use it: the transcript_check-unverifiable clause only appears when the loaded evals.json actually declares a transcript_check assertion (resolve now runs before preflight — it is read-only, so nothing is built before preflight can still reject).
• Missing dispatch recipes now warn: a harness without a [dispatch] exec_template (OpenCode, exec-less BYOH descriptors) gets a preflight warning naming the generic-handoff fallback, via a new has_dispatch_recipes() trait probe.
• Fixed a latent inconsistency: --guard --no-stage previously kept guard=true through artifact rendering (manifest/runbook carried guard args for an unguarded run); the guard is now resolved before rendering.

Before/after (bare eval-magic run, claude-code)

Before: run is unguarded unless the operator remembered --guard.

After:

Preparing my-skill iteration-1 (new-skill)
  ...
  guard: armed — the write guard blocks out-of-env writes during dispatch (--no-guard to opt out)

🛡 Write guard armed: a PreToolUse hook is staged in .claude/settings.local.json

Bare eval-magic run --harness opencode now messages every limitation:

⚠ --harness opencode declares no write guard — … Pass --no-guard to acknowledge and silence this.
⚠ --harness opencode declares no transcript parser — tokens/duration go unrecorded and …
⚠ --harness opencode declares no dispatch exec recipe — RUNBOOK.md and dispatch-manifest.md carry handoff guidance …

Docs

run --help (--guard/--no-guard/--no-stage/--harness), top-level --help examples, README (quick-start examples, Cost & confirmation, Harnesses fallback paragraph), docs/progressive-enhancements.md, docs/byoh.md, per-harness notes, harnesses/template.toml.

Verification

• cargo test — 637 tests green (9 new unit tests incl. the full guard decision table; 9 new/renamed integration tests across lifecycle/codex/opencode/byoh/basics)
• cargo fmt --check, cargo clippy --all-targets -- -D warnings
• End-to-end drives: bare run auto-arms (plan line + banner + marker/hook in each env); --no-guard installs nothing and silences the warning; --guard --no-guard is a parse error; opencode/BYOH degrade with the warnings above

🤖 Generated with Claude Code