Stream large engine output over SSE instead of truncating at 2 MB

[pavlof01]

Context

Engine output is capped at MAX_OUTPUT_BYTES (2 MB) in packages/engine-runtime/src/run.ts; past that the run is flagged meta.outputTruncated and the tail is discarded. Flags like --print-all-code, --trace-ignition and --print-maglev-graphs blow past 2 MB on small snippets, so exactly the most interesting V8 output is the output the user cannot see. The user also waits for the whole run before seeing a single line.

Proposal

• Stream the engine's stdout/stderr to the gateway and on to the client over SSE (or chunked text/event-stream), with the existing cap raised or removed for streamed responses.
• Keep the current JSON response as the default so the API contract and the cache are unchanged; streaming is opt-in per request.
• The output pane appends as chunks arrive, with a hard client-side cap on rendered lines (virtualise or paginate) so the browser survives a 50 MB dump.
• Do not cache streamed responses; document that.

Acceptance criteria

• A --print-all-code run shows its first lines before the run completes.
• A run that exceeds the old cap is fully retrievable.
• Non-streaming clients see no behaviour change.