feat(schema): body nodes and parameters carry their id in analysis.json

[rahlk]

Closes #164. Parity with codeanalyzer-python 1.4.1's #180 (design in its #176).

Python now stamps an id on every body node in analysis.json — <callable-id>@<local>, the same value its Neo4j projection merges the node on — and on every parameter — <callable-id>@formal_in:<i>, the L4 formal_in vertex that carries it, emitted as a forward reference below level 4. The shared vocabulary.md promises both. TypeScript emitted neither.

What this adds

• TSBodyNode.id on every body node at every level, and TSCallableParameter.id on every parameter at every level
• One definition of the join key. The composition rule lived in two private fq copies (build/neo4j/project.ts and dataflow/attach.ts). globalOrdinal in schema/ids.ts is now the single implementation; both delegate to it.
• stampBodyIds runs after each body emitter — at the end of the L1 emitter per callable, and once over every callable at the end of applyDataflow. The second is deliberately not per-callee: L4 writes actual_in/actual_out vertices into caller bodies, so a per-callee stamp would miss them.

Verification

• Two agreement tests, mirroring python's, pin the property against the other projection rather than a hand-typed string: every body node's id equals the set of projected :TSBodyNode keys for that callable; every parameters[i].id equals the @formal_in:i vertex's id, with the vertex's of naming the parameter. Plus an L1 test that call nodes and parameters carry ids before any dataflow exists.
• Mutation-checked: skipping the body stamp and changing the @ separator each fail the tests
• bun test green; container tests run (the projection's fq now delegates); schema.neo4j.json unchanged — :TSBodyNode already carried id as its merge key and parameters aren't graph nodes
• SCHEMA_VERSION stays 2.0.0 (chore(schema)!: collapse the schema version back to 2.0.0 #144); decision recorded in .claude/SCHEMA_DECISIONS.md

Note on the test's shape

Because both projections now share one definition, an agreement test alone can't detect a defect in globalOrdinal — both sides would move together. The L4 test therefore also pins the literal @formal_in:<i> shape, and the L3 test asserts a call node's id is never its callee. That's the same trade python made.