perf(runtime): bound Cattle artifact restore latency and memory

[Yevanchen]

Context

PR #559 restores the latest ready artifacts when a recycled Cattle Session gets a fresh workspace. The current path restores files serially: each R2 object is fully buffered as a Uint8Array, base64-encoded, and then written through the Sandbox RPC.

This prioritizes correctness, but large files increase Worker peak memory and encoding/GC pressure, while many files add cold-start latency linearly. This is a follow-up optimization, not a blocker for #559.

Related: #559, #387, langgenius/mosoo-agent-driver#117

Goal

Measure and reduce artifact rehydration latency and Worker memory without weakening Session isolation, latest-per-source-path selection, audit history, original paths, or fail-closed handling for missing objects.

Scope

• Add observability for artifact count, total bytes, largest file, R2 read time, encoding/write time, end-to-end restore time, and failures.
• Establish a repeatable baseline for 1/10/50/100 files with mixed sizes, including p50/p95 and peak memory.
• Based on measurements, evaluate the smallest safe improvement: bounded concurrency, fewer byte/base64 copies, or streaming if supported by the Sandbox API.
• Define safe per-file and per-restore byte limits with an actionable error; do not introduce unbounded concurrency.

Acceptance criteria

• A repeatable benchmark or integration test compares before/after behavior.
• Restore latency no longer grows strictly through fully serial file processing, or measurements justify a different optimization.
• Peak memory has a documented bound and all artifacts are never buffered simultaneously.
• Existing Cattle continuation, cross-Session isolation, latest-version, and missing-object tests remain green.
• Production rollout and rollback signals identify whether time is spent in R2 read, encoding, or Sandbox write.