fix/neo4j: lossless field, import & call-graph projection (#156, #157, #158) + relative-path build fix

[rahlk]

Fixes #156, #157, #158, all in the --emit neo4j projection path, plus a coupled build-path papercut found while reproducing #158.

#156 — every JField collapses into one node

projectField keyed the node id by Field.getName(), which is always null for a field declaration (the IR keys fields by their variables list). All fields of a class collapsed to a single <fqn>#field#null node under the j_field_id unique constraint. Now keyed by the joined variable names with a positional fallback.

• daytrader8: #field#null 122 → 0; distinct JField ids → 339.

#157 — imports lose the type name

Single-type imports were stripped to a :JPackage, and multiple imports from one package collapsed to one edge. Now non-wildcard, non-static imports link to a :JType (materializing an external ghost when it isn't a project type); wildcard/static-member imports keep :JPackage; and the full path rides on every J_IMPORTS edge. Adds path to the J_IMPORTS schema and regenerates schema.neo4j.json.

• daytrader8: J_IMPORTS → JType 0 → 1470, → JPackage 771 → 9 (just wildcards).

#158 — J_CALLS undercount

The call-graph vertex signature rewrites constructors from <init>(...) to the simple class name for readability, while the :JCallable node id keeps the raw <init>(...) signature — so every constructor call edge was gated out. Key J_CALLS endpoints off callable_declaration, which preserves the symbol-table signature verbatim (verified: zero ambiguity conflicts).

• daytrader8: J_CALLS 1562 → 1715 (~97% of call-graph edges with both endpoints present). The remainder are synthetic/implicit callables with no source node (implicit default ctors, access$ accessors, bridge methods, anonymous classes) — correctly dropped.

Note: the "287 edges / 14%" reported in #158 against 2.4.0 does not reproduce on current HEAD — projection was already at ~75%; this recovers the constructor edges to ~97%.

Relative-path build fix (found while reproducing #158)

• A relative --input doubled the project path: the build runs with its working dir at the project root, but gradleBuild passed a relative -p (maven was already absolutized, gradle wasn't), resolving to <root>/<root> → build fails → 0 application classes. Normalized to absolute once at the build entry point.
• That failure was then swallowed: buildProjectAndStreamClassFiles returns an empty list (never null) on failure, but ScopeUtils only guarded == null, so analysis proceeded with an empty scope and died with a cryptic WALA "Could not create entrypoint callsites". Now guards emptiness and reports the real cause.

Tests

• New GraphProjectorCallGraphTest (constructor resolution, gating of unresolved targets, signature fallback).
• Neo4jSchemaConformanceTest green (covers --emit neo4j: all JField nodes share id <fqn>#field#null, collapsing every field of a class into one node #156/--emit neo4j: imports lose the type name (J_IMPORTS only links to JPackage, never JType) #157 + regenerated schema).
• All three fixes verified end-to-end on the daytrader8 fixture at --analysis-level 2.
• Docker-gated tests (Neo4jBoltWriterTest, CodeAnalyzerIntegrationTest) not run locally — exercised in CI.