fix(cyberbattlesim): reconcile SDL evidence requirements, simulator capture, and backend manifest

[Brad-Edwards]

Objective

Reconcile the authored CyberBattleSim SDL/task evidence requirements, the data the pinned simulator can actually expose, the data needed for equivalence testing, and the adapter's implemented capture and backend manifest.

Scope

• Inventory every data-capture requirement in the CyberBattleSim SDL and task and map it to the pinned native simulator source.
• Capture a complete per-step attacker action/outcome log sufficient to trace each episode and investigate outliers.
• Capture the availability series and distinguish simulator termination, truncation/cutoff, and other terminal causes as far as the native source permits.
• Capture and report stochastic controls and RNG bindings honestly; disclose every unbound source of randomness.
• Ensure the resulting artifacts are sufficient to recompute the declared equivalence measures from preserved per-run data.
• Correct authored requirements that overclaim simulator capability and explicitly disclose unavailable or lossy data.
• Extend the adapter capture path and make its backend manifest truthful at field/artifact granularity.
• Remove the false attacker-action-log satisfaction claim and every other hardcoded evidence-reference claim in this path.

Acceptance criteria

• Every SDL/task capture requirement maps to a native source, an implemented adapter capture path, a manifest declaration, and an emitted portable artifact field.
• Each episode can be traced through its attacker actions and outcomes to availability, termination/cutoff cause, reward, and equivalence inputs.
• Availability and termination-cause comparisons can be recomputed from preserved evidence without trusting a summary claim.
• RNG/control coverage and limitations are explicit and machine-readable.
• Every unavailable or lossy requirement is explicit and cannot be admitted as satisfied.
• Tests exercise the complete requirement-to-capability-to-artifact chain and fail on missing or falsely claimed data.
• The corrected equivalence experiment is rerunnable from the preserved protocol and emits the required evidence.
• The canonical verification graph passes.

References

• test(cyberbattlesim): reproduce the selected public baseline through RAES #30

[doublewhy]

Status: blocked during the naming and breaking-change freeze.

Do not implement or publish contract, schema, manifest, runtime, package, or CLI changes from this issue yet. Resume requires both the pending naming decision and the field/artifact capture-capability work in OpenRAE/rae#1112. Static inventories or regression-fixture design may be prepared only if they preserve current production behavior.