Normalize conversation-loop result shape at the Agents API boundary

[chubes4]

Problem

conversation-loop.php normalizes the Agents API conversation result, then adds Data Machine-specific fields directly at the top level.

Code reference on origin/main (008878b6):

• inc/Engine/AI/conversation-loop.php:300-342 — adds completed, last_tool_calls, runtime_tool_pending, runtime_tool_pending_requests, completion assertion fields, max_turns_reached, and warning on the top-level result.

Related closed result-shape/context issues include #1514, #1569, #1667, and #2020.

Why this matters

Some of these fields are useful diagnostics, but the contract mixes canonical Agents API result fields with Data Machine runtime/provenance fields. For 1.0, callers should not need to know which top-level keys are substrate contract and which are product/runtime adornments.

Acceptance criteria

• Define the canonical Data Machine conversation result boundary for 1.0.
• Keep Agents API result fields canonical at top level, or explicitly document Data Machine's wrapper contract.
• Move Data Machine-only diagnostics into a namespaced key such as runtime_provenance, metadata.datamachine, or another agreed container.
• Preserve required chat/UI behavior without relying on legacy aliases like max_turns_reached unless intentionally retained.
• Tests cover budget exceeded, runtime tool pending, completion assertions, and chat response shaping.

AI assistance

• AI assistance: Yes
• Tool(s): OpenCode (GPT-5.5)
• Used for: 1.0 technical-debt audit and issue drafting; Chris requested tracking issues for the findings.