fix: Preserve unread exec output when max_output_tokens truncates a chunk #144

Summary

exec_runtime drops output data when a response chunk is truncated by max_output_tokens.

• Affected code: /Users/mcell/Desktop/workspace/memo-cli/packages/tools/src/tools/exec_runtime.ts:279
• Related code: /Users/mcell/Desktop/workspace/memo-cli/packages/tools/src/tools/exec_runtime.ts:281

Why this matters

Tool output can be silently lost, which breaks correctness for long outputs and makes follow-up write_stdin polling incomplete.

Repro

1. Start exec_command with a command that emits large output and keeps running briefly.
2. Set small max_output_tokens (e.g. 10).
3. First response shows Original token count much larger than returned output (expected truncation).
4. Poll same session with write_stdin and no input.
5. Observe Original token count: 0 and empty output, instead of remaining unread tail.

I confirmed locally with:

• cmd: node -e "process.stdout.write('X'.repeat(5000)); setTimeout(() => {}, 4000)"
• First chunk Original token count: 1250 but only 40 chars returned.
• Second poll returned nothing.

Expected

If output is truncated, unread remainder should still be available in subsequent polls.

Actual

readOffset is advanced to full delta length before truncation remainder is retained, so tail is discarded.

Suggested fix

• Advance readOffset only by delivered bytes/chars, not full delta.
• Keep a per-session pending-output buffer (or ring buffer) for remainder.
• Add regression test covering truncation + subsequent poll.

Requested implementer

Please let memo handle this fix.