resolver: note razor marker removals; make retention counters mean what they log

[pbednarcik]

Follow-up to the #531 review — the four non-blocking notes. (The blocking one and the
test-comment fix I see you already landed in 9c9f8c9, along with the CI guards in
1d4c9d9 — thanks, that saved me the biggest half of this list.)

1. Razor marker removals now dirty the retention. The real adjacency mutation in
   razor_using.go is the RemoveEdge consuming the @using markers, and it had no
   removal helper — the note above it was inert, exactly as the review said. Added
   noteImportEdgeRemovals (removal-side sibling of noteImportEdgeReindexes: known
   file → per-file dirty, provenance-less → wholesale generation), wired it to the
   marker consumption, and dropped the inert call. The new test follows the pattern
   your pin test established: it drives the production resolveRazorUsings pass with
   no note* call in the test body, and the projected adjacency keeps the marker
   target unresolved so the file can only be served by the adjacency retention —
   watched it fail without the fix (projection calls = 1, want 2). The write-site
   audit table now covers RemoveEdge sites too and lists razor under removals;
   csharp_partial_merge.go's removals are classified kind-bounded (implements/extends
   moves between type nodes — imports edges never originate at a type).

2. Counters now mean what they log. reach_fallback is a zap.Bool (it was a
   flag logged as a count), and reach_retained is renamed reach_occupancy — it
   reports the retention map's size, which is what it always measured; per-page
   retention hits were already reach_adj_cached. One log key changes name, so
   anything parsing the prepare line needs the new key.

3. Identity-preserving skip: kept, re-commented. The comment now names the one
   shape that actually produces it — the deferred-LSP terminal clears, where
   terminality is meta-only so OldTo equals the target — and says it runs after the
   pass's final prepare. I kept the skip rather than removing it: it encodes a true
   invariant (an identity-preserving rewrite cannot change stored adjacency), and if a
   future writer produces the shape mid-pass, skipping is the correct behavior where
   removal would silently evict the files the retention serves.

4. Cap comment matches what the cap now bounds — both retentions, with the
   adjacency entries called out as the heavy ones (~5 MB real, tens of MB at the cap,
   per your measurement).

Testing: the new test failed before the fix and passes after; ./internal/resolver
and ./internal/indexer FAIL name-sets on Windows are identical to clean main
(zero new, zero repaired); gofmt clean.