Nice, focused change. The core insight—that getLastNVisualLines and the expanded-content normalization were being computed unconditionally even when hidden (neither is shown) or expanded (preview isn't shown)—is legitimate and the fix in thinking.tsx preserves the original render output for all three states (I traced showPreview/showFull through the new branches and they match the old semantics: hidden computes neither, expanded skips the preview-line wrap, preview skips the expanded-markdown normalization).

The thinking-block.tsx fast-path for blocks.length === 1 is a reasonable micro-optimization to avoid a throwaway array from .map() on every streaming token, though its actual impact is probably marginal compared to the wrapping fix — that's fine, it's still correct and harmless.

Good to see unit tests added for all three collapse states (preview, hidden, expanded) verifying the rendered frame contents rather than just internal state, which is the right level to test a terminal UI component at.

A couple of things worth double checking before this lands: (1) confirm getLastNVisualLines and expandedContent truly have no side effects relied on elsewhere (e.g. some parent measuring width off lines) — a quick grep of call sites would settle this. (2) The benchmark numbers in the PR description are useful context but aren't part of the diff/test suite, so they're not verifiable in review — that's fine for a description but the correctness argument stands independently.

Small, in-scope, testable — worth porting.