Apply specified types in place on a batched working copy

[ondrejmirtes]

Applying a SpecifiedTypes batch derived one whole scope — a full holder-map copy plus scope construction — per specification, and one more per array-dim level of the parent-intersection cascade inside specifyExpressionType(). This PR restructures the apply side to write in place:

• specifyExpressionType() opens a single unpublished copy (openSpecificationScope()) and writes holders straight into its maps (specifyExpressionTypeInPlace()). Every lazily-derived view (resolvedTypes, truthy/falsey scopes, the promoted-native scope) is reset after each write, so every read answers exactly as it did on a per-specification fresh scope.
• The application loop batches all consecutive specifications into one working copy, inlining the addTypeToExpression()/removeTypeFromExpression() math; operations that go through other scope derivations (isset certainty, overwriting assignment) publish it and the next specification opens a fresh one. The conditional-expressions tail moves to processConditionalExpressionsAfterSpecifying().
• The type-specification list is now built inline at its only caller, with alternative-form entries sorted together with sure/sure-not entries — restoring the shortest-expression-first guarantee (parents before children) for them too. ScopeOps::buildTypeSpecifications() and its native mirror are removed (smoke differential and signature parity green).
• filterBySpecifiedTypes() is renamed to applySpecifiedTypes() — the method does not filter, it applies computed narrowing. It is public but not @api.
• The isset-certainty path keeps the held type instead of re-reading it via getType(), which broadens Maybe-certainty holders and would overwrite a co-applied narrowing.

Zero expectation churn: full test suite green in both turbo modes (17844 tests), self-analysis and CS clean, turbo smoke differential + signature parity green.

Benchmark note: the local machine is currently running a backup; I will post a commit-toggled ABBA user-CPU comparison as a comment once it quiets down.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DaBZjgksga4c5s6Q9FniY7

[ondrejmirtes]

Benchmark (machine now quiet): commit-toggled ABBA, same tree, no-turbo, bin/phpstan analyse src/Analyser src/Rules with cleared result cache, user CPU, 6 pairs.

• base (2.2.x): median 49.63s — [48.63, 49.32, 49.37, 49.90, 50.26, 51.17]
• this PR: median 49.73s — [47.49, 48.58, 48.82, 50.64, 50.80, 51.10]
• delta +0.19% median; paired mean diff −0.20s (sd 2.16s)

Neutral within this run's noise floor — the value here is convergence (and the removed native mirror is a 1–3 element build+sort per application, so its deletion is micro). Happy to run a longer full self-analysis ABBA or a turbo-mode pair if wanted.