refactor: modularize parallel runtime and harden safety

[dickymoore]

This is the third PR in the parallel execution stack. It is opened as a draft to avoid accidental merge before the earlier stack entries land.

Please review/merge order:

1. fix: codex loop robustness #2: Codex bug fixes.
2. feat: add parallel story execution MVP #4: parallel execution MVP.
3. feat: add parallel runtime controls and recovery #5: runtime controls and recovery.
4. This PR: safety, review semantics, and modularization cleanup.

Scope:

• Replaces source result.env worker result loading with explicit allowlisted key/value parsing.
• Uses private mktemp -d script snapshot directories.
• Updates code-review gate return codes so 0 means clean/success and nonzero means follow-up/error.
• Simplifies worker wait state so wait_for_worker_completion no longer takes nine nameref parameters.
• Splits the large core into sourced modules:
  • lib/runtime.sh
  • lib/review-loop.sh
  • lib/parallel.sh
• Updates wrappers and installer so modules load correctly from repo and installed layouts.

Tested:

• bash -n on core, wrappers, installer, and all lib modules.
• Direct source ./ralph-loop-core.sh smoke test proving sourced functions are available.
• Wrapper snapshot --help smoke test.
• Simulated installed-layout smoke test using ralph-loop-lib.
• Worker result parser rejection test for an unexpected/malicious result key.
• Fake Codex parallel smoke test with RALPH_CONCURRENCY=2 in a temporary git project: two ready-for-dev stories ran through worker worktrees, fake dev/code-review workflows completed, controller integrated both worker outputs, and both stories ended as done in sprint status.

No live Claude/Codex agent workflow was invoked; the provider was faked to avoid spending real model calls during branch validation.