perf(sidecar): raw-binary readFileSync transfer (skip base64)

Background

Guest binary fs.readFileSync(path) (no encoding) goes
_fs.readFileBinary → _fsReadFileBinary → mapped to fs.readFileSync with the
binary flag (crates/execution/src/v8_runtime.rs:183) → host returns the bytes as
a base64-encoded JSON object (javascript_sync_rpc_bytes_value:
{ "__agentOsType": "bytes", "base64": "..." }), and the guest decodes with
Buffer.from(base64, "base64") (crates/execution/assets/v8-bridge.source.js).

So every binary file read pays base64 encode (host) + decode (guest) plus ~33%
payload inflation. The base64 round-trip is wasted work.

Proposal

Transfer file bytes as raw binary instead of base64. The bridge response
protocol already has a raw-binary path (status = 2 / Uint8Array, see
crates/v8-runtime/src/host_call.rs). Route readFileSync's binary result through
that path and have the guest do Buffer.from(uint8array) directly.

Why it wasn't done in PR #77

The sync-RPC return type is a serde_json::Value, which can't carry raw bytes —
hence the base64 representation. Switching readFileSync to the status=2
raw-binary response is a cross-cutting protocol change that affects every
binary file read, so it was deferred rather than rushed (read was already 6.2×
faster from the pump fix in #77). Needs careful implementation + tests for:
binary readFile (sync + promises), writeFileBinary symmetry, and the
{__agentOsType:"bytes"} consumers elsewhere.

Expected impact

Removes base64 encode/decode + ~33% payload inflation per binary read; helps
read-heavy guests (bundlers, asset pipelines, large node_modules).

Start with crates/execution/src/v8_runtime.rs:183, crates/execution/assets/v8-bridge.source.js, and crates/v8-runtime/src/host_call.rs to trace the existing binary response paths. Inspect the other {__agentOsType:"bytes"} consumers, then add coverage for binary sync and promise reads plus writeFileBinary symmetry; done means binary reads use the raw-binary path without base64.