Play Framework

spike(testrunner): classify DB tests reached via interface dispatch (PF-154)

Root cause of the downstream (JClaw) autotest flake tax: usesDatabase()'s
ASM reachability walk follows only STATIC references, so a unit test that
touches the DB through dynamic dispatch on an interface — where the
concrete impl isn't statically reachable from the test's call graph
(e.g. MemoryStoreFactory.get() typed as the MemoryStore interface, impl
JpaMemoryStore which references models.Memory + play.db.jpa.JPA) — is
mis-tagged U: (pure/parallel). It then runs on the 16-permit parallel
lane and its Fixtures/JPA activity races the serialized D: lane on the
shared jdbc:h2:mem DB, wiping rows mid-run.

Fix: on an interface method call (visitMethodInsn isInterface=true),
enqueue every application class implementing that interface via
Play.classes.getAssignableClasses, so the walk reaches the concrete
DB-touching impl and classifies D: correctly. Keeps the parallelism win
for genuinely-pure tests; visited-set dedups the impls.

PROTOTYPE — API-verified against fork source but NOT yet built/measured.
Validation (PF-154 AC): build the fork, point a JClaw worktree's
composite build at it, dump the U:/D: classification to name the
currently-misclassified test(s), and measure flake reduction + suite-time
delta before any fork-main merge.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>