fix: allow fully covered selected-dimension reviews

[eminos]

Summary

A review that explicitly selects 18 dimensions and assesses all 18 is currently rejected because two other language dimensions are absent from the run. The same gate blocks recovery with review --import-run.

 trusted import coverage
-  block if any language dimension is absent
+  block if any selected dimension is absent

Restrict the missing-dimension gate and its error detail to the packet selection. The broader scorecard coverage notice remains informational.

Evidence

• Before: three new synthetic regressions fail: a complete 18-dimension selection is blocked, error detail includes an unselected dimension, and completed-run replay is blocked. The full-selection fail-closed regression already passes.
  After: all four pass; normal replay preserves previously trusted unselected assessment values, dates, and metadata without --allow-partial or launching review batches.
• Current upstream branch: 50 focused tests passed, Ruff and diff checks passed.
• Full suite on the local integration base b1f6010: 5901 passed / 70 failed / 4 skipped, versus 5897 / 70 / 4 on that exact clean base. All 70 failed IDs are identical; the sandbox cannot download required tree-sitter parser artifacts. This comparison is limited to that integration base, not a full-suite claim about current upstream.
• Normal CLI replay of retained completed outputs also succeeded in isolated disposable state. No active project state or original outputs were changed; no review batches were rerun.

Merge Danger

Door: two-way

Blast Radius: imports

Trusted imports still fail when a selected dimension is unassessed. Full/default selections retain their existing fail-closed behavior. The change does not remove or replace unselected assessments.