perf(core): cut peak live memory 17.3% (taffy capacity, MeasureCtx dedupe)

[quake]

Summary

Cuts pocketjs-core peak live memory 1,429,840 -> 1,182,457 bytes (-17.3%) on the new deterministic membench journey, with zero behavior change (drawlist checksum byte-identical, 127 tests pass, wasm32 no_std build clean).

• Drop the unread MeasureCtx.text duplicate of every text run
• Rebuild the taffy tree at subtree-sized capacity, freeing the old maps first (slotmap geometric growth left ~160 dead slots x ~1 KB)
• Slim MeasureCtx to the shaped size; drop per-level children clones in collect_subtree/collect_run

engine/core/examples/membench.rs is the new counting-allocator benchmark (boot -> steady -> churn -> burst; prints PEAK_LIVE_BYTES).

Allocation-churn trade-off (measured)

Structural rebuilds now re-allocate the taffy maps instead of reusing historical capacity:

	baseline	head
TOTAL_ALLOC_BYTES	11,416,332	14,755,184
ALLOC_COUNT	109,081	101,314 (−7.1%)

The +3.3 MB of total allocation bytes is a handful of large blocks per structural rebuild (recycled by the PSP arena's O(1) size-class free lists); the allocation count goes down.

Benchmark integrity

NODES reports the true live node count via a harness-side hierarchy model mirroring Ui::destroy_node's subtree semantics (a call-counting counter drifted by +35 after churn). DRAWLIST_CHECKSUM is byte-identical across baseline and head.

Note on CI

Workflow runs on this fork PR require maintainer approval (action_required); the numbers above are local runs of cargo test, cargo run --example membench, and cargo build --target wasm32-unknown-unknown, reproducible with the commands in the diff.

[HalfSweet]

Thanks for the optimization work. Removing the duplicated data from MeasureCtx and avoiding children.clone() during recursive traversal both look reasonable.

My main concern is that the current benchmark measures live bytes using the host allocator, while the actual PSP constraint is the arena high-water mark. The benchmark also includes some allocations from the test harness itself, so the current results are not enough to show that rebuilding and resizing the Taffy tree on every structural relayout is a net improvement on PSP.

There are also a few unrelated or incomplete benchmark artifacts in this PR, along with some inconsistencies between the benchmark code and its comments. I suggest splitting the work: merge the low-risk deduplication and clone removal first, then fix the benchmark and provide PPSSPP/PSP arena_bump_bytes and layout timing results before evaluating the Taffy storage rebuild strategy separately.

For now, I’m requesting changes.

[quake]

I have narrowed this PR to the two low-risk core changes:

• keep the MeasureCtx data deduplication
• keep the recursive traversal changes that avoid children.clone()
• remove the Taffy capacity-rebuild strategy
• remove the unrelated/incomplete benchmark artifacts

The Taffy storage rebuild will be proposed in a separate PR after adding PSP/PPSSPP arena_bump_bytes and layout timing measurements. This PR now passes the core test suite: 127 tests passed.