highlight.highlight() runs a real lexer walk over the source. The same
code block re-renders many times in normal use — a streaming update
on the *next* block still re-runs the parent transcript, the React
StrictMode double-invoke in dev re-renders every block, and the
code-block toolbar's hover-reveal re-renders the block's wrapper.

Add a 200-entry LRU keyed on a djb2 hash of (lang + code). The cache
stores the original source alongside the rendered HTML and verifies
the source matches before serving — a hash collision (vanishingly
unlikely with djb2 on multi-KB code) returns a cache miss instead of
the wrong highlight. Map iteration is insertion-order, so the
oldest entry is at the head; eviction is a single delete().

Size 200 covers the visible viewport (~30 blocks) plus an overshoot
for the transcript's re-render window. Memory is bounded: each entry
is the (code, html) pair, so worst case is ~200 * average-block-size
bytes; in practice 50KB total for a busy transcript.

Public API is unchanged. tsc --noEmit passes.