Stop reporting a failed Codex integration as a healthy host

[MongLong0214]

Reproduced, then fixed. The defect is mine, introduced in #698 while fixing
#697.

detail   "Codex registration added and live-verified; plugin step failed"
healthy  true
ok       true
exit     0

#698 appended the plugin outcome to detail and left healthy alone. ok is
computed from the field, not the sentence — so the installer exited 0 on a
requested integration that did not happen.

What changed

Codex is two requested integrations, and the host is healthy only if both are.
The plugin outcome is a value now, not a sentence, and composing it with the MCP
result is a named function so the composition can be tested directly. Reaching
it through a spawned codex would have meant no test at all.

An unhealthy registration stays unhealthy whatever the plugin did — without that,
a successful plugin step would resurrect a host whose registration is broken.

Verified

Two mutations, two different failures:

healthy: plugin.ok → mcp.healthy   → "reports unhealthy when the plugin step fails"
remove the unhealthy early return  → "never turns an unhealthy registration healthy"

Which half worked stays readable in detail: the registration may be usable
while the plugin is not, and a reader who cannot tell them apart repairs the
wrong one.

66 tests passing across the installer files. Canonical artifact ca608927.

This is the same shape as everything found today — a check that names a failure
and does not enforce it — and I wrote it yesterday.

[github-actions]

CommitLore — record lint

Trailers: clean — 3 commits in origin/main..e1c13d045c082506b53b7ca920318670f86c91e9
Active constraints: not read — commitlore: git log --follow accepts exactly one pathspec, so renames are not followed for 8 paths; query one path at a time to follow its rename chain (8 changed paths)

_{Trailer violations fail this check. Active constraints are informational — they are what the repository already decided, not a verdict on this PR.}