Teach agents to validate calls before fan-out

[yuelgrace1810-ops]

Problem

In two independent bounded graph-search trajectories, the agent sent eight parallel graph.search.atlas calls with the same invalid limit: 2000 payload. The contract allows at most 100, so one unvalidated payload mistake became eight identical errors.

Solution

Teach the managed Jacobian math Skill to run one representative canary call before repeating a payload shape in parallel, and fan out only after that call succeeds. Keep the repository and packaged Skill byte-identical and within the existing 4 KiB budget.

Evaluation evidence

• Eager routing trajectory: eight duplicated invalid Atlas calls.
• Selective routing trajectory: eight duplicated invalid Atlas calls.
• Focused forward test with this patch: one successful order-0 canary preceded the batch; all later Atlas calls used valid limits; zero INVALID_REQUEST diagnostics.

Overlap

PR #786 documents exact contract inspection. It does not address parallel amplification of one unvalidated payload. No open PR or current Skill guidance contains an equivalent canary-before-fan-out rule.

Testing

• uv run pytest tests/unit/tooling/test_codex_visibility.py -q — 14 passed
• skill-creator quick_validate.py — passed
• Ruff — passed
• managed Skill parity — passed
• git diff --check — passed

Trust & Compatibility Impact

Instruction-only change. No mathematical capability, verifier, artifact format, public API, or assurance behavior changes.

Checklist

• Relevant focused tests pass
• Packaged and repository Skills remain identical
• Skill remains within the 4 KiB budget

[morluto]

this should be at the tool layer not the skill (try asking your codex to inspect the validaiton, how it processes the tools)