impact_analysis can answer SAFE or "read by 0 models" when it can't know #1667

Found during the v0.12.9 release review (#1647).

Reproduced on the built 0.12.9 binary with edited jaffle-shop manifests:

• Leaf model with an exposure: titled Impact: SAFE with "Change is safe to make" while listing the dashboard (impact-analysis.ts severity and safe line count only downstream models).
• Column that doesn't exist, or is quoted or padded: "amount", no_such_col read as "read by 0 models", with no "column not found".
• Semantic model or metric reading the column, with no exposure: not reported, so it reads as 0 and safe.
• Stale manifest: generated_at and the manifest path appear nowhere, and an out-of-date compiled_code gives a confident 0.
• Name shared by versions or packages: findModel takes the first match, so stg_v can resolve to v1 where ref('stg_v') means the latest, giving a false SAFE. A unique_id works.

Suggested: no SAFE when exposures exist; normalize and check the column; add a one-line note that semantic models and metrics aren't checked; report the manifest path and age, and warn on stale; list other candidates for an ambiguous name.