feat: always-on Insights — correlation analysis over viewing habits

Conservative, live insights (ADR-0010) derived from each user's diary:
rating lift by genre (first-time watches only), genre seasonality vs a
uniform 1/12 month baseline, medium weekend-lift, and watch-time rhythm.

- domain: WatchRow model, InsightsQuery port, pure insights_analyzer with
  all honesty thresholds (sample minimums, strength cutoffs, suppressions
  for import noise), unit-tested
- adapters: thin Sqlite/Postgres fetch twins, reusing WrapUp's genre fetch
- application: get_insights use case, InsightsGroup in Deps
- composition/server: wiring through Repositories/DatabaseOutput
- presentation: GET /api/v1/insights handler, typed serde-tagged DTOs,
  openapi doc
- spa: /insights route (shadcn cards), insights feature hook, profile
  entry link, en locale strings
- server test: end-to-end seed sqlite → HTTP → rating lift + rhythm

Also records ADR-0010, defines "Insight" in the glossary, and includes
pre-existing worktree comment-stripping cleanups.

crates/adapters/metadata/src/tmdb.rs

@@ -68,7 +68,7 @@ impl TmdbProvider {
         struct Details {
             imdb_id: Option<String>,
             title: String,
-            release_date: String, // "YYYY-MM-DD"
+            release_date: String,
             poster_path: Option<String>,
             credits: Credits,
         }
@@ -88,7 +88,6 @@ impl TmdbProvider {
                 ))
             })?;
 
-        // Prefer IMDB ID; fall back to "tmdb:{id}" so the record is still usable.
         let raw_id = d
             .imdb_id
             .filter(|s| !s.is_empty())