Consume unchanged effect-free subtrees during convergence passes

[ondrejmirtes]

Third extraction chunk from the single-pass analyser branch (follow-up to #6249).

Convergence passes re-walked the whole loop body every round. Now each pass chains the previous pass's storage as its baseline and keeps the full ExpressionResult of every expression it walks in a per-site frame: an expression whose stored result is effect-free (no scope derived, no throw/impure points) and whose read state is unchanged (variables and tracked expression holders, both attribute-cached per node) is consumed instead of re-walked — only the loop-carried parts of the body re-walk.

To keep #6249's replay firing on consuming passes, every walked result tags its emission segment [recording, start, end) and a consumption splices that segment from the pass that last walked the subtree into the consuming pass's recording; only a segment-less consumption gaps the pass and clears its replay candidacy.

One subtlety: the old-world pricing walks that resolveType() starts for short-circuit operators re-enter processExprNode() mid-pass with a fresh throwaway storage. They emit nowhere, so they may consume stored results but must not splice, store, or tag — keying those operations on the pass's own storage identity fixed emission duplication the naive version produced.

On PHPStan's own handler-heavy code, ~29% of pass expression walks are consumed. The CPU effect is modest — the expensive subtrees (calls with throw/impure points) fail the effect-free gate by construction, so what is consumable is also cheap to re-walk:

corpus	before	after	Δ
self-analysis (--debug src, 3+3 interleaved)	80.30s	79.30s	−1.2%
WordPress core (e2e corpus, level 0)	~173s	~172s	~0 to −2%, noise-limited

Analysis output is byte-identical on full self-analysis with the feature on vs off; full test suite, make phpstan, and CS are green. No turbo-ext changes (the per-site frames are plain PHP state; the shadowed storage class is untouched).

🤖 Generated with Claude Code

https://claude.ai/code/session_01GnwgpaeUXRkgSDyg95tfK8

[ondrejmirtes]

Closing unmerged: A/B measurements show the consume mechanism buys at most ~1% anywhere — upstream (−1.2% self-analysis, ~0 on WordPress) and on the single-pass branch (0 to −1%, noise-limited, point estimate even slightly negative). The expensive subtrees fail the effect-free gate by construction, so what is consumable is also cheap to re-walk, and the gate itself costs comparably. The replay (#6249) stays as the convergence win; this complexity does not pay for itself.