Isolated GLM reasoning-boundary diagnostic

One request on October 25 local time (October 26 UTC), preserved as negative functional evidence. No performance ranking, promotion, Ignis run, or engine patch.

• Plan: bounded controls and conditional second case; execution and live-identity flags reset to false; original authorized packet remains private.
• Source freeze, runtime identity, source registry, projection ledger.
• Original runner result, token observations, offline usage replay.
• Summary, closure, friction, role manifest.

The native result correctly leaves engine drain unknown. The separate post-run owner capture proves sampled drain. The original usage warning remains in place; its offline correction does not change the failed output. Exact request/response and frozen dirty-source patch remain protected private originals. This public bundle was assembled after the diagnostic, not before it; pre-request plan, source freeze and execution gate were retained privately before dispatch.