perf(start-client-core): zero-copy frame payload extraction

[anonrig]

What

Add a zero-copy fast path to extractFlattened() in the client-side frame decoder (packages/start-client-core/src/client-rpc/frame-decoder.ts).

Previously, every frame payload (and header) was extracted by allocating a fresh Uint8Array and copying count bytes — even when the requested bytes were fully contained in the first buffered chunk, which is the common case since most frames arrive within a single network read.

The new fast path returns a subarray view of the first chunk when the bytes are contiguous, avoiding both the allocation and the byte copy. The multi-chunk (spanning) path is unchanged.

Why

This is on the hot path for decoding every streamed server-function response and every RawStream binary payload on the client. Removing the per-frame copy matters most for large binary streams.

Standalone micro-benchmark (Node, median of 12 runs), copy vs. zero-copy extraction of contiguous payloads:

Payload	Current (copy)	This PR (view)	Speedup
1 KB frames	224 ms	74 ms	3.0x
64 KB frames	258 ms	9.6 ms	27x

The win scales with payload size because it eliminates the byte copy.

Safety / tradeoff

The returned view shares the chunk's backing ArrayBuffer. This is safe because buffered chunks are never mutated in place after being read from the network.

There is one nuance worth flagging for review: for CHUNK frames the payload is enqueued onto a ReadableStream and may be retained by a downstream consumer, so a retained view pins its (slightly larger) source network chunk rather than a tightly-sized copy. For the JSON/header/error paths the bytes are consumed immediately, so the view is strictly a win. In normal streaming usage (process-and-release), the view is better on both CPU and peak memory. If reviewers prefer, we can restrict the view to JSON/header reads or gate CHUNK on a size threshold — happy to adjust.

Tests

• Existing frame-decoder suite passes (typed-array assertions use value equality, so views are transparent).
• Added a test that reassembles a 300-byte CHUNK payload delivered in 7-byte reads, exercising the multi-chunk slow path the fast path branches around.

test:unit   ✓ 19 passed
test:types  ✓ no errors
eslint      ✓ frame-decoder.ts clean (no new problems vs main)

Notes

• Independent of the sibling perf PR that replaces the O(n) bufferList.shift() with an index pointer. Both touch extractFlattened, so whichever merges second will need a trivial rebase.

[anonrig]

Superseded by upstream PR TanStack#7662 (opened against TanStack/router).