fix: persist bounded execute_code result bodies

[Tom-StrategicAutomation]

Summary

• Persist a normalized result record in each completed execute_code audit manifest.
• Bound the stored value to 32 KiB (32768 bytes), reusing the executor's existing PREVIEW_BYTES budget so audit bodies cannot grow without limit.
• Serialize completion and abandonment with the existing per-file lock and atomic JSON writer, preventing concurrent terminal updates from overwriting one another.

The additive manifest structure is:

{
  "result": {
    "parsed_as_json": true,
    "representation_type": "json",
    "original_size_bytes": 123,
    "stored_size_bytes": 101,
    "max_stored_size_bytes": 32768,
    "truncated": false,
    "value": "..."
  }
}

JSON values are canonicalized with sorted keys and compact, ASCII-safe encoding. Non-JSON output remains text. Redaction runs before deterministic UTF-8 truncation, and no unbounded raw result is copied elsewhere in the manifest.

Redaction policy

• Recursively copy dictionaries and lists without mutating the caller's result object.
• Redact case-insensitive named fields after hyphen/whitespace normalization: password, passwd, secret, token, access_token, refresh_token, api_key, apikey, authorization, cookie, set-cookie, and private_key.
• Apply the same explicit field policy to plain or malformed text, including quoted field names and :/= assignments.
• Redact complete or partial PEM private-key blocks, Authorization/Bearer values, Cookie/Set-Cookie values, JWTs, and common sk-, GitHub gh*_, and Slack xox*- token forms.

Integrity

• Source digest, audit ID, and command identity remain byte-for-byte stable from intent to completion.
• Receipt digest/status/command validation happens before result inspection and remains fail-closed for mismatched intent receipts.
• Pending batch reconciliation is unchanged.
• History continues to contain only the manifest reference; result bodies are not duplicated there.
• Completion/completion and completion/abandonment races are serialized under the receipt lock; writes remain atomic.

Tests

• Focused: uv run pytest -q tests/guard/gates/test_terminal_policy.py -k execute_code — 29 passed, 66 deselected in 4.31s.
• Full post-commit suite: uv run pytest — 496 passed in 57.58s.
• uv run ruff check . — All checks passed.
• uv run ruff format --check . — 106 files already formatted.
• uv run python scripts/violin_guard.py check-release — all release checks passed, including its internal Ruff and test-suite gates.
• git diff --check and git diff --check dev...HEAD — passed.

All Python verification ran on Linux in the official ghcr.io/astral-sh/uv:python3.11-trixie-slim container.

Closes #12