fix(l4): join the SDG port lattice to the statement ddg

[rahlk]

Closes #227

What

SdgPortEdges adds the four edge classes that were missing between the port lattice and the statement graph — the same shape codeanalyzer-python#115 and codeanalyzer-typescript#80/#81 landed:

edge	meaning
@formal_in:k → use	mirrors the @entry definition of parameter k
return → @formal_out	the statements that hand a value back
statement → <call>/actual_in:i	the statement that evaluates the arguments
<call>/actual_out → statement	the statement that consumes the result

Each is a re-rooting of something the statement ddg already says, so the pass is pure and engine-free. It runs after SummaryPass, so summaries are still computed from the statement-level ddg alone.

A call's own body node carries no dependence edges, so actuals attach to the innermost statement containing the call, or to the call itself for a bare-call statement (foo(x);, where statement and call share an addressing anchor). The attachment is deliberately coarse — every actual of a site is fed by the one statement containing it, not by the reaching definition of that particular argument. L4 may add reach, never drop it, and the precise binding is recoverable from the statement's own incoming edges.

No contract change

No new node kind, edge type or field; prov stays the closed {ssa, points-to} enum. The graph contract holds at 2.0.0.

Verified — daytrader8 at -a 4

	before	after
ddg edges touching a port	0	5,108
formal_in with out-degree 0	1,176 of 1,176	140 of 1,176
total ddg	5,475	10,583

The 140 that remain are parameters no statement reads — investmentReturn, orderCompleted and pingTwoPhase are single-throw bodies, and removeHolding never reads orderID.

Output validates against analysis.v2.schema.json; two runs are byte-identical; every new endpoint is a body node of its own callable, so all 5,108 reach the Neo4j projection.

./gradlew test: 514/515 pass, 4 skipped. The one failure is CodeAnalyzerIntegrationTest ("Could not find a valid Docker environment"), which fails identically on main here.

Known gap

Only the AST L3 engine roots parameter flow at @entry; --l3-engine wala keeps NormalStatement endpoints only, so it has no such edge to mirror and its formal_in ports stay unattached. The other three classes are engine-independent.

Propagation

• codeanalyzer-python / codeanalyzer-typescript — none: both already landed this shape; Java was the last of the three.
• codeanalyzer-schema — none: emission only, contract unchanged.
• python-sdk — follow-on when its codeanalyzer-java pin moves to the release carrying this: the port-edge probe passes on its own, so the gate on slice_forward/paths_between/flows_to_call/flows_to_argument can lift with no SDK code change, and any committed Java -a 4 fixtures need regenerating (ddg roughly doubles).
• Stacking note: with fix(ddg): drop dependence edges whose endpoint is not a body node #229 (issue ddg edges reference endpoints that are not body nodes of their own callable #228) merged, the daytrader8 total is 10,493 rather than 10,583 — the 90 dangling points-to edges that PR drops.