feat(ids): a module's can:// id names the module, not where it was checked out

[rahlk]

Spec: 2026-09-17-manifest-derived-module-prefix.md · Tracking: codellm-devkit/.github#92 · Spec review: codellm-devkit/.github#91

A module id gains a segment naming the module as its own manifest declares it, between the language segment and the path within that module:

canjv -i /path/to/daytrader-web-service --app-name daytrader

before  can://daytrader/java/src/main/java/org/apache/.../RunStatsDataBean.java
after   can://daytrader/java/daytrader-web-service/src/main/java/org/apache/.../RunStatsDataBean.java

No staging directory, no -f, and no pointing --input at a parent that sweeps in sibling repos.

The coordinate comes from the manifest rather than the directory name so it survives a rename, a relocated checkout, or a differently-laid-out CI workspace. The gate fixture demonstrates the difference: its directory is mvnw-working-test but its pom declares my-javaee-mvn, and the id carries the latter.

Two rules that are load-bearing, not incidental

The symbol_table key does not change. It stays the real --input-relative path. Uniqueness is the key's job and a path is unique by construction where an artifactId is not; with no path field on the module object, the key is also the only thing that still locates a file on disk. Consequence: the key is no longer the tail of the id. L3WalaOverlays recovered the application id by stripping the key off a module id, so it is now passed the real one and that derivation is deleted.

A contested coordinate is applied to neither claimant. The id is the join key for call_graph endpoints, so a coordinate claimed by two module directories would merge two distinct callables onto one node. Vendored duplicates are exactly where this bites — several services each building one shared internal library under the same module name. Verified:

[WARN] Module coordinate 'daytrader-core' is declared by 2 modules (svc-a/lib, svc-b/lib);
       ids for these keep their plain relative path, since a shared coordinate would collide them

can://dt/java/svc-a/lib/src/main/java/com/shared/Bean.java
can://dt/java/svc-b/lib/src/main/java/com/shared/Bean.java

Without that rule the change would regress "ids are distinct by construction" — the property it was chosen for.

A module with no readable coordinate keeps exactly the id it had before, so a loose source tree is unaffected: can://bare/java/src/main/java/com/foo/Bare.java.

Tests

16 new, written failing first: ModuleCoordinatesTest (7), ModulePrefixesTest (7), L1BuildContextTest (2). They pin the parent/dependency artifactId distinction, the DOCTYPE refusal, nearest-enclosing-manifest for reactor children, the --input search bound, the contested-coordinate fallback, and the key/id separation.

L3WalaOverlaysTest is deleted — its only test reflected into the derivation this PR removes. L2CallGraphGateTest's expected ids are updated; that fixture declares rootProject.name = 'call-graph-test', so it is also a live example of the documented app-name/coordinate redundancy (can://call-graph-test/java/call-graph-test/...), which the spec §5 accepts rather than special-cases.

Verification

Gate	Result
./gradlew test	619 tests, 1 failed — the pre-existing CodeAnalyzerIntegrationTest Testcontainers/Docker initializationError
Schema conformance	-a 1/2/3/4 all validate against analysis.v2.schema.json
Monotonicity	json(-a 1) ⊆ (-a 2) ⊆ (-a 3) ⊆ (-a 4)
Determinism	two -a 4 --eager runs byte-identical

Reviewer note: the warning is invisible by default

Log.toConsole gates every level behind isVerbose() (Log.java:128), and --verbose defaults to false, so without -v the analyzer prints no WARN and no ERROR at all. The contested-coordinate warning above only appears under -v.

That is pre-existing and wider than this PR — it also makes the README's "it degrades and says so on stderr" guidance false by default — but it means this PR's warning does not currently reach anyone who does not already pass -v. Worth fixing separately; flagging it rather than silently relying on a warning nobody sees.

Does not fix

#269 — a single analysis spanning source roots that declare the same fully-qualified type binds every caller to the last root, silently. This makes a bundled tree addressable, not correctly resolved.