Precision loss after a branch merges with dead code #2129

Split off from #627 / #2128.

-AignoreDeadCode (#2128) stops errors from being reported on code that is physically
inside a dead branch (a literal-condition if, or code dataflow proves is CFG-unreachable). It
does not fix a related but distinct problem: precision loss on code after a branch, caused
by merging with a dead alternative. Example:

@Nullable Object x;
if (true) {
    x = new Object();
} else {
    x = null;
}
x.toString();  // false positive today, even though the else branch is dead

CFAbstractStore.leastUpperBound runs at the join point during GenericAnnotatedTypeFactory .analyze(), a dataflow pass that is entirely separate from, and runs before, whatever the
type-checking visitor later chooses to check. Skipping visitation of the dead branch (what
#2128 does) has no effect on that pass, so x still gets widened to @Nullable at the
join and x.toString() is still flagged.

#101 ("Introduce "bottom store" to reduce false positive in dead branch") fixes exactly
this, by making the untaken branch of a literal-condition if/while/for/ternary produce a
lattice-bottom store instead of a normal one, so leastUpperBound with it is the identity and
the live branch's precision is preserved across the merge. That PR is not mergeable as-is
(unconditional — no option gate, changes default behavior for everyone; touches dataflow/'s
Store.FlowRule and CFGTranslationPhaseOne for ternary handling; requires new constructors on
every checker's Store/Analysis pair using names that no longer match current code, e.g.
InitializationStore no longer exists under that name), but the core technique — an isBottom
flag on CFAbstractStore, a bottom-store singleton on CFAbstractAnalysis, and a
CFAbstractTransfer.visitBooleanLiteral hook for if/while/for conditions only (dropping
#101's ternary/FlowRule changes as the riskiest, least-reviewed part) — is worth
reviving, gated by -AignoreDeadCode, with a safe createEmptyStore() fallback for any
checker's Store that doesn't override the new bottom-store method (so nothing else needs to
change to keep compiling).

Neither #101 nor #633's own tests actually exercised this join-point case (both only
tested errors inside a dead branch), so this was scoped out of #2128 rather than bundled
in.

🤖 Generated with Claude Code