Fix regular trace ATIF trajectory shape

[mmprotest]

Motivation

• Ensure the regular-runner trace-debug trajectory artifact (trajectory.json) conforms to ATIF v1.7 while keeping the new trace-debug artifact feature intact.
• Preserve existing provider-reported input/output token telemetry in the full transcript artifact and avoid changing runtime gating or export failure behavior.

Description

• Edited villani_code/atif_export.py to emit an ATIF v1.7-shaped trajectory with sequential step_id, source and message fields, ATIF-standard per-agent metrics.prompt_tokens/metrics.completion_tokens, llm_call_count, tool_calls entries using tool_call_id/function_name/arguments, and observation.results entries with source_call_id and serialized content; redaction is preserved.
• Kept transcript.full.json generation unchanged except for ensuring tests still validate provider-reported usage telemetry is present and unmodified.
• Updated tests/test_regular_trace_artifacts.py to assert the ATIF v1.7 root/step/tool/observation/final-metrics shape and to validate telemetry preservation, redaction, gating, and non-fatal export behavior.
• Made minimal, targeted changes only and preserved the existing integration boundary and artifact-writing guards.

Testing

• Ran pytest -q tests/test_regular_trace_artifacts.py tests/test_debug_mode_and_recorder.py which passed (22 passed).
• Ran pytest -q tests/test_trace_runtime_path.py which passed (8 passed).
• Ran pytest -q tests/test_loop.py and observed pre-existing failures unrelated to this patch (3 failing tests due to an UnboundLocalError in Runner.run retry path); these failures were present after the change and the patch did not modify that codepath.
• Harbor validator was not available in the environment so no Harbor schema run was performed (harbor not installed).
• Confirmations: the patch modified villani_code/atif_export.py and tests/test_regular_trace_artifacts.py; trace-debug transcript token telemetry remains preserved in transcript.full.json; trajectory.json now uses ATIF v1.7 field names and structures as required; and benchmark, autonomous/Villani, small-model, planning-only, prompts, tool flow, validation, and existing transcript/gating behavior were not changed and remain covered by the tests.

---

Codex Task