agent enable: a failed package install is no longer a success

[graywolf336]

atomic agent enable falls back to built-in hooks when the integration package can't be installed. The fallback is right — turns still get recorded — but it exited 0 and printed the success blurb after the warning, so a partial install looked like a whole one.

What's actually missing in that state is the agent's Atomic skills and system prompt. Hook recording still works, which is why the gap survives unnoticed: nothing errors, runs complete, and the agent is quietly less capable than the same agent on the next machine. That is how it went unnoticed for weeks in practice.

Before — exit 0:

⚠ Integration install for Codex failed: … — falling back to built-in hooks.
✓ Installed 6 hooks for Codex
Each agent turn will be recorded as an Atomic change with: …

After — exit 4:

⚠ Partial install — the integration package did not land:
    Codex: cannot read /nonexistent/atomic-integration.toml …

  Turns are still recorded by the built-in hooks, but these agents
  are missing their Atomic skills and system prompt.
  Retry with 'atomic agent enable --force' once the package is reachable,
  or install from a local checkout with '--from <path>'.
✗ Remote operation failed: integration package install failed for Codex — …

RemoteError, not Internal: an unreachable package isn't a bug in Atomic, and Internal tells the user to file an issue.

Behaviour change. Anything scripting this will now fail on a partial install where it previously passed. That's the point; || true opts out where the fallback is genuinely acceptable. Flagging it explicitly in case that's too blunt for CI somewhere — happy to put it behind a flag instead.

Verified against a real repo: healthy install still exits 0, a forced package failure exits 4 with the block above. Plus 3 unit tests on the error itself, cargo fmt --all --check and clippy -- -D warnings clean, 11 tests in agent::enable pass.