fix(ids): declaration merging mints one id per facet; both facets survive in types{} (#177)

[rahlk]

Closes #177. Spec: docs/design/specs/declaration-merging-ids.md (D1–D5, mirrored in .claude/SCHEMA_DECISIONS.md). Analyzer 1.5.0; Neo4j contract stays 2.0.0.

What was wrong

Every facet of one name has the same signature, and ids were minted from signatures alone, so a value and a type of one name (const TableOption = () => … + interface TableOption) shared an id and --emit neo4j MERGEd them onto one node with both labels and the last writer's kind. Worse, a type/type merge (class C + interface C) was "later kind wins" in types{} — the class vanished from analysis.json entirely.

What changes

• Value facets keep the bare id. A type whose bare id is already minted by a callable or field in the same scope becomes <id>#type. Callables anchor call edges, body-node ids, @formal_in:N and parameters[i].id; a type's id is only referenced through extends_ids/implements_ids.
• Type/type merges keep both facets. Builder order (class → interface → enum → alias → namespace): the first keeps the bare key and id; each later facet is keyed Name#<kind> in types{} and gets <id>#<kind> (#interface, #namespace, …).
• Only on collision; signatures never move. A merge-free project is byte-identical (verified: sample-app -a 4 output on this branch is byte-identical to main). idBySig keeps the value facet for call-graph re-identification, callee backfill and homing; a new typeIdBySig gives heritage the type facet; a split is not an L1 collision.
• Heritage resolution in the builder now picks the first type-like declaration of a merged symbol instead of declaration [0] (which was the value), so implements TableOption resolves to the interface facet.
• Neo4j: nothing to change — distinct ids are distinct nodes; kind and labels agree on every node.

Tests

test/declaration-merging.test.ts: value+interface, alias+arrow, alias+const, class+interface, function+namespace — distinct ids per facet, both facets in types{}, implements → #type facet, extends → bare class, no L1 collision, member ids under the bare class id, and the projection has no node with two kind labels. bun test: 345 pass.

No sibling precedent: python rebinds on a later def, Java forbids the construct.