Config-read dataflow tiers: close non-literal keys over the L3 DDG and the L4 call graph #236

Plan (optional)

Spec: codellm-devkit/.github → docs/design/specs/2026-09-07-java-config-reads-and-entrypoint-report.md (decision D30, tiers). Epic: codellm-devkit/.github#62. Builds on #232 / PR #233 — branch from that branch, not main.

Problem

#232 shipped the literal tier: a config read resolves when its key argument is written as a string literal at the call site. Everything else lands in config_reads_unresolved with reason="non-literal", which on a real codebase is the common shape:

String key = "spring.datasource.url";
env.getProperty(key);                  // non-literal today

String read(String name) { return System.getenv(name); }   // non-literal today

codeanalyzer-python closes both over its dataflow graph (its #162 Task 3), so a Java graph reports strictly fewer readers than a Python one for the same code shape.

Scope boundary

In scope: the two dataflow tiers over the existing L3/L4 overlays, and the prov bookkeeping that goes with them.

Out of scope: any new dataflow machinery — both tiers read the DDG and call graph the analyzer already emits. Also out: the entrypoint report (#234), python-sdk accessors (epic child 4, deferred), CRUD (#187).

Goals

• L3 intra tier. A key argument that is a bare name closes when every DDG-reaching definition of that name at the call site is the same single string literal. Filter prov.contains("ssa"); require the DDG use node's span to contain the call node's span; accept only a single-target <name> = "literal" definition (VariableDeclarationExpr with one declarator, or AssignExpr onto a NameExpr).
• L4 interproc tier. A key that names a parameter of its enclosing callable closes when the parameter is never locally redefined and every call site targeting that callable supplies the same literal — directly, or through one hop of caller-side intra closure.
• prov: ["dataflow"] on tier-produced edges; unresolved records carry every tier attempted, so ["literal", "dataflow"] at L3+.
• Tier gating: intra at -a 3+, interproc at -a 4+. A read resolved at a lower tier is never recomputed.
• ConfigUses takes the analysis level, module source, and the L2 call graph it now needs.

Caveats and known risks

• The ssa filter is what keeps -a 3 ⊆ -a 4 true, and it is not optional. At -a 4, L4WalaOverlays.java:201 adds points-to to the prov of existing edges and mints alias-widened ones. An alias edge can connect this variable's use to an unrelated write that is not a name = "literal" shape; the shape check then refuses it, and — since any non-closing reaching def kills resolution — that refusal would remove an edge the ssa-only L3 set resolved cleanly. Monotonicity broken by a widening, which is the exact failure the additive contract forbids. codeanalyzer-python documents hitting this.
• Any non-closing reaching def must kill the resolution, not be skipped. Two paths assigning different literals means the read is genuinely ambiguous, and picking one would be a confident wrong answer.
• The span-containment check is load-bearing. The CFG/DDG is statement-level while a call body node is keyed by its own narrower span, so a def's recorded use site is the enclosing statement. Matching by containment covers both the bare-expression-statement case and the common return env.getProperty(key); nesting without special-casing either.
• SdgPortEdges also emits prov:["ssa"] ddg edges, whose endpoints are synthetic @formal_in/@actual_out vertices with no span. The span guard rejects them; do not relax it.
• The interproc tier needs completeness of the call-site set. A callable reachable through a call edge the analyzer did not resolve must not close, or a caller supplying a different key is silently ignored.
• A read the intra tier closed to a literal that matches no declared key must not be re-derived from callers by the interproc tier: the local rebinding is what the callee actually reads, and the caller's argument would misattribute.

Definition of done

• On a fixture carrying String k = "server.port"; env.getProperty(k);, -a 1 reports it non-literal and -a 3 resolves it with prov: ["dataflow"].
• On a fixture whose key is a parameter with two callers passing the same literal, -a 4 resolves it; with two callers passing different literals, it stays unresolved at every level.
• A local reassigned to two different literals on two paths stays unresolved — no arbitrary pick.
• Monotonicity on a fixture that carries a literal read, an intra-closable read and an interproc-closable read: config_uses(-a 1) ⊆ config_uses(-a 3) ⊆ config_uses(-a 4), checked as set containment of (src, dst) pairs, and the literal-tier edges keep prov: ["literal"] at every level.
• Running -a 4 with --l3-engine wala (which adds points-to prov) resolves the same set as -a 3 plus interproc — never fewer.
• analysis.v2.schema.json still validates a -a 4 payload; the L1/L3/L4 conformance gates stay green.