review-graph: compile successful multiline validation commands without losing identity #151

Problem

The review-graph planner and worker-payload publication accept a validation command containing real newlines, but compile-node rejects the successful execution with:

review_graph_runtime: command must be one non-empty line

Observed while reviewing acgetchell/la-stack#275 at 6decbf8e30c5e151738ea148caa8d10797a3a285, using dotfiles 9b2077b. All ten package-validation commands exited zero, but the seventh contained a multiline Python program passed to uv run --locked ... python -I -c. Compilation failed, leaving validation reconciliation and both synthesis nodes blocked.

Cause

agents/.agents/skills/review-graph/scripts/review_graph_runtime.py:2052 renders executions using _required_text(raw, "command"). That helper rejects newline characters. The native execution verifier in review_graph_plan.py also compares the displayed command directly to the raw dispatched command.

Reproduction

1. Plan/materialize a validation unit containing a shell command such as:

   uv run --locked python -c 'print("first")
   print("second")'

2. Run it successfully and publish its exact command, exit code 0, and timing in a ValidationPayload.
3. Compile the node using matching captures and workspace snapshots.
4. Observe the single-line text error despite accepted planning/publication and successful execution.

No validator rerun should be necessary to recover the existing evidence.

Acceptance criteria

• Safely render multiline commands on one native-report field line while retaining the exact raw command in payloads and command identity digests.
• Update native verification consistently; altered commands must still fail exact-identity checks.
• Preserve compatibility with existing single-line evidence.
• Cover LF, CRLF, quotes/backslashes, leading/trailing whitespace, and command text resembling report headings or execution records.
• Add regression coverage through payload publication, compilation and verification.
• Recompile already successful evidence without changing the dispatched command or rerunning validation.

A draft implementation was explored and then set aside; it is not committed. The full affected test suite was interrupted when deferring this work, so it is not validated as a completed fix.