Persist bounded output/result manifests for audited execute_code calls

[Dan-StrategicAutomation]

Summary

The v2.0.0 execute_code audit contract requires engagement metadata, saves the Python source, and records a digest/status/duration in history. It does not preserve the tool result or output in a structured receipt comparable to violin_exec.

Affected paths on dev at dc53bbb:

• plugins/violin_guard/code_execution_audit.py
• plugins/violin_guard/__init__.py
• plugins/violin_guard/execution.py
• tests/guard/test_terminal_policy.py

Current behavior

record_completion(source, result, duration_ms) writes only:

• evidence/<phase>/execute-code-<timestamp>-<digest>.py
• one history row containing digest, target, duration, and status

_result_summary() parses enough of result to decide success/error, then discards the result body.

Impact

Complex scripts used for structured HTTP requests or payload construction are source-auditable but outcome-poor. Reviewers cannot reconstruct what the script printed or returned from the same receipt model used by guarded command execution.

Suggested fix

Write a bounded JSON sidecar manifest containing:

• schema version and receipt ID;
• source digest and relative source path;
• engagement, phase, target, and session metadata;
• start/completion or duration fields;
• normalized status/error information;
• a size-capped result/output field;
• truncation and redaction indicators.

Reference the JSON manifest from history. Preserve the source file separately. Avoid storing unbounded results or secrets.

Acceptance criteria

• Successful, failed, non-JSON, and oversized results produce valid manifests.
• Output truncation is explicit and deterministic.
• Sensitive-field redaction follows repository privacy conventions.
• History references the JSON manifest used for review.
• Existing source-digest validation remains unchanged.
• Regression tests prove the result body can be audited without reading tool chat output.