Follow-up: precision/completeness backlog from per-task reviews

[rahlk]

From the v1 whole-branch review (epic #1); ledger-triaged. Stitcher: use wiring as tiebreaker for ambiguous client evidence; unanchor on equal-score server-anchor ties (currently dict-order wins DERIVED). Protos: nested messages register as pkg.Inner not pkg.Outer.Inner (FQN collision risk). Blast: comment the depth-cap Pareto corner (strong-at-cap label can shadow shallow weak path). Demo: align _naive_tokens skip set with detect._SKIP_DIRS (vendor/dist/build inflate the headline denominator); _fetch success-path unit test; per-process .partial suffix (concurrent demo race). Facts: Java-side dropped-edge test. Datastore: key: node attrs collision on dedup. Detect: root-vs-src same-name dedup test; verify *.sln-only C# services. Spec: sql-column/Column extraction (descoped from v1).

[rahlk]

From final-review verification: topology.py docstring overstates compounding — when a HOSTS edge wins, the HANDLES edge's provenance is never consulted (and setdefault keeps only the first HANDLES edge). Unreachable today (stitch emits HANDLES only as DERIVED-STATIC); align docstring or consult both when addressing this backlog.