Add ATIF v1.7 and full transcript artifacts for regular TRACE debug runs

[mmprotest]

Motivation

• Provide two additional debug artifacts for regular (non-Villani) TRACE-mode runs: a full transcript (transcript.full.json) and an ATIF v1.7 trajectory (trajectory.json).
• Keep the change minimal and non-invasive by reusing existing in-memory transcript/message captures and the repository’s redaction behavior, without changing runtime behaviour, prompts, tool logic, benchmarking, or autonomous flows.

Description

• New exporter module villani_code/atif_export.py with pure, side-effect-free builders build_full_transcript_artifact(...) and build_atif_trajectory(...) that serialize existing transcript and messages data and reuse maybe_redact_payload/token-normalization logic. Redaction is applied consistently when requested.
• DebugRecorder.write_regular_trace_artifacts(...) added to call the pure exporters and safely write transcript.full.json and trajectory.json into the existing debug run directory using the recorder’s safe-write isolation.
• Runner (villani_code/state.py) gains a single narrow eligibility predicate _should_write_regular_trace_artifacts() and terminal calls to _write_regular_trace_artifacts(...) at the existing finalisation points for regular-run completions/failures; this guard enforces: TRACE mode, runtime execution (not planning), not small-model, not villani/autonomous, not benchmark-enabled, and presence of a debug recorder.
• Tests: added tests/test_regular_trace_artifacts.py with focused cases validating eligibility rules, ATIF structure and ordering, tool-call ↔ tool-result linkage, token metrics aggregation, non-duplication of cumulative requests in ATIF steps, redaction application, non-fatal export errors, and that execution results remain unchanged except for the presence of the two new files.

Testing

• Code compiled: python -m compileall -q villani_code tests/test_regular_trace_artifacts.py (succeeded).
• Focused unit tests for the new behaviour: pytest -q tests/test_regular_trace_artifacts.py — 12 passed (all tests in that file passed).
• Larger subset run: pytest -q tests/test_regular_trace_artifacts.py tests/test_debug_mode_and_recorder.py tests/test_trace_runtime_path.py tests/test_trace_summary.py tests/test_state_runtime.py — 87 passed across that set.
• Full test suite: pytest -q — the run showed unrelated pre-existing failures in other areas (11 failing, 607 passed, 11 skipped); the failing tests are outside the scope of this change (examples observed: tests/test_hardening_pass.py::test_interactive_plan_approval_still_works, tests/test_interactive_tokens.py::test_interactive_shell_is_textual_wrapper, several tests/test_loop.py cases, tests/test_mission_state_runtime.py::test_autonomous_summary_mirrors_to_mission_state, planning/TUI contract tests, and tests/test_villani_hygiene.py::test_low_authority_paths_are_blocked_from_mutation).

Notes / confirmations

• Exact files added/modified: villani_code/atif_export.py (new), villani_code/debug_recorder.py (edit), villani_code/state.py (edit), and tests/test_regular_trace_artifacts.py (new test file).
• The change does not modify benchmark, autonomous/Villani, small-model production codepaths, prompts, CLI/config surfaces, tool execution logic, approvals, retries, timeouts, validation, termination logic, message-history construction, or existing transcript generation; the only externally visible change is that eligible regular TRACE debug runs now write transcript.full.json and trajectory.json into the debug run directory.
• ATIF fields intentionally left absent/null if not captured by the current runner: provider may be null when not supplied, model_name may be null if unavailable, step-level timestamps/durations/costs are not fabricated, and message fields may be null when the corresponding response has no text blocks.

---

Codex Task