Retain full Claude and Codex red-team evidence for Apptainer

[merlerm]

The existing agent probes did not provide a complete retained Apptainer run across both backends, and some verdicts confused echoed markers or broken requests with evidence.

Add a coordinator that runs every catalog case, keeps per-case raw streams/scripts/broker logs, continues after failures, and distinguishes failed, inconclusive, and passed results. Add strict interpreter/path/source inventories and a rendered-policy import attack. Randomize undisclosed canaries, recognize Codex tool events, require genuine result markers, and verify raw strict commands with fresh working connections and explicit denials.

Validation: 7 coordinator/checker tests passed at this layer. Both original full suites completed (44 cases plus network audit): Codex 38 passed / 6 inconclusive / 1 failed; Claude 35 / 8 / 2. Review found real package/bytecode defects fixed by earlier stack layers, checker false positives, and unresolved refusals/budget exhaustion. Targeted reruns and the final deterministic/network audits are documented; the original runs are not represented as all green. The new rendered-policy case expands the catalog to 45 plus network audit.

Paid probes are opt-in. Raw local audit logs, images, credentials, and unrelated campaign files are not committed.

Stack integration validation: 258 tests passed, 4 image/runtime tests skipped in the isolated worktree; pylint passed and mypy passed for 11 changed modules. Paid cluster audit results are summarized above where relevant.

_{Stack created with GitHub Stacks CLI}

[merlerm]

Consolidated into #230. Retained evidence and full-suite execution now live in the original red_team_sandbox.py; the separate coordinator has been removed.