fix(sourcehunt): bound machine progress records

[ropoctl]

Summary

• enforce the machine protocol 1 MiB outgoing-record ceiling
• project SourceHunt progress onto bounded metadata and counts instead of transporting bulk file, symbol, and finding-ID arrays
• retain bounded terminal compaction as a final safety net
• report staged early exits accurately instead of labeling deep runs as quick depth

Checkpoints remain workspace artifacts and do not cross the machine protocol.

Why

Machine-protocol records must remain bounded regardless of repository size. SourceHunt could complete preprocessing successfully and then fail while emitting an oversized progress or result record across FD3.

Progress is observability metadata, not artifact transport. This change makes that boundary explicit and keeps progress and terminal records within the protocol ceiling.

Validation

• 49 passed, 1 deselected in focused machine and checkpoint tests
• Ruff passed
• git diff --check passed
• live deep jq preprocess completed with a 378,808-byte filesystem checkpoint, a 1,291-byte maximum machine record, and no checkpoint in the terminal payload
• verified depth remained deep and the staged-exit detail was Run stopped after preprocess

Open PR overlap

Reviewed the open SourceHunt PRs that touch runner.py or sourcehunt.py. They cover hunter progress limits, batch sizing, trace persistence, sandbox setup, and hunt/resume behavior; none owns the machine-record boundary or progress-event projection.