Fold rebuild reassignments into initializers and reclaim final locals

[steve-aom-elliott]

Fixes #155.

Two related improvements to MigrateMapperSettersToBuilder for mappers configured in a helper method away from their construction site.

1. Fold mapper = mapper.rebuild()...build(); into the initializer when adjacent

When a rebuild reassignment is the statement immediately following the declaration, fold the two into a single-statement form. The issue #155 shape:

final ObjectMapper mapper = initObjectMapper();
mapper.disable(SerializationFeature.WRITE_DATES_AS_TIMESTAMPS);
mapper.disable(SerializationFeature.WRITE_DATE_TIMESTAMPS_AS_NANOSECONDS);
return mapper;

now migrates to:

final JsonMapper mapper = initObjectMapper().rebuild()
        .disable(SerializationFeature.WRITE_DATES_AS_TIMESTAMPS)
        .disable(SerializationFeature.WRITE_DATE_TIMESTAMPS_AS_NANOSECONDS)
        .build();
return mapper;

The fold intentionally requires the reassignment to be the very next statement — any intervening read of the local (as in the existing trailingSetterAfterGapOnLocal shape, where System.out.println(mapper) sits between the setters) prevents the fold and preserves observable semantics.

2. Reclaim final locals so the rebuild rewrite applies

Before this PR, final locals were unconditionally skipped by the rebuild-assignment path added in #153 — the receiver has to be reassignable — so #155's exact shape (final ObjectMapper mapper = ...; followed by two configuration calls) got the TODO fallback and shipped broken code once the SerializationFeature → DateTimeFeature rename ran. Now:

• If the fold above succeeds, the declaration is initialized once and final is preserved.
• If the fold can't apply (an intervening read forces the two-statement form), the recipe strips the final modifier so the rebuild reassignment is legal.
• Un-finalizing is skipped when the local is referenced inside any lambda or anonymous-class body in its scope, because a subsequent reassignment would break Java's effective-final capture requirement. Those cases keep the existing TODO.

How

• foldRebuildIntoInitializer post-pass: walks each block, and when it sees T name = <init>; immediately followed by name = name.rebuild()....build();, merges them into T name = <init>.rebuild()....build();. Uses the existing RebuildParts / tryParseRebuildAssignment helpers to reconstruct the chain rooted at the initializer.
• foldRebuildIntoInitializer runs after coalesceRebuildAssignments (so consecutive rebuild reassignments become one chain before folding) and before unfinalizeDeclarations. All three post-passes are now scheduled from visitBlock so ordering is explicit; visitMethodInvocation records reclaim candidates in a per-visitor Set<UUID> instead of scheduling doAfterVisit directly.
• unfinalizeDeclarations inspects each candidate declaration and only strips final when the initializer doesn't already contain a .rebuild() chain — i.e. only when the fold couldn't consume the reassignment.
• findLocalDeclaration is factored out of isReassignableLocal and reused by reclaimableFinalLocalDeclarationId. isReferencedInsideCapture walks the enclosing method body for lambda/anonymous-class capture; those decls stay final and get the TODO.
• Final fields intentionally still get the TODO — this change is scoped to locals.

Test plan

• ./gradlew test — full suite passes, including five new cases under RewriteAsRebuild:
  • finalLocalIsUnfinalizedAndRewritten — final local with an intervening read: two-statement form, final stripped.
  • finalLocalFromMethodCallCoalescesRebuildChain — the issue Jackson 2 to 3 migration leaves broken code when ObjectMapper is created in one method and configured in another #155 reproducer: coalesce + fold + final preserved.
  • nonFinalLocalFromMethodCallFoldsIntoInitializer — proves the fold applies to non-final locals too.
  • finalLocalCapturedByLambdaKeepsTodoComment — capture guard for Supplier<JsonMapper> lambdas.
  • finalLocalCapturedByAnonymousClassKeepsTodoComment — capture guard for anonymous inner classes.
• Existing trailingSetterAfterGapOnLocal still asserts the two-statement form (intervening read prevents fold).
• Existing finalFieldStillGetsTodoComment unchanged: final fields intentionally excluded.

[steve-aom-elliott]

Trimming comments now

[timtebeek]

Nice to see this expanded once again, thanks!