Handle oversized WASM host responses without misleading fetch errors #16

Problem

The WASM SDK uses a fixed 1 MiB mallocBuffer for host-managed allocations. A fetch response's JSON-encoded envelope must fit in the space remaining after the invocation request and prior host allocations. On allocation failure the host yields an empty response, and the SDK reports plugin: fetch: empty response from host; provider code can misreport this as network unreachability.

This is not hypothetical: a read-only GitHub REST kubernetes/kubernetes issue-list page with per_page=100 was 1,168,331 bytes; the SDK fetch envelope was 1,238,310 bytes. A scratch WASM harness using the pinned SDK reproduced the failed allocation and empty-fetch error; a small control succeeded. With a 41-byte request, an envelope of 1,048,535 bytes succeeded and 1,048,536 failed.

Requested outcome

• Handle large host fetches and outgoing results without silently exhausting a shared allocation arena; retain explicit resource limits.
• Return a clear actionable size-limit error rather than an empty response or misleading network error when a response cannot be delivered.
• Cover request-plus-host-allocation boundaries and oversized responses in tests using the actual WASM runtime and host path.
• Audit sibling host allocation calls for the same failure mode and keep existing plugin compatibility.

The proposed GitHub issue-list implementation (Paca-AI/paca-plugin-github#33) temporarily uses one issue per page to reduce exposure, but this is not a universal fix: one response can still be oversized, and prior allocations consume arena space. Companion SDK contract PR: #15.