fix: include sidecar eval tokens in optimizer totals

[latentwill]

Summary

Optimizer and eval token displays now count sidecar-backed trader decisions instead of showing only the mutator/judge slice. Eval reports fall back to eval_runs.actual_*_tokens when no linked model_calls exist, cycle summaries add those linked fallback actuals to the cycle meter, and cycle exports include per-experiment token counts in both JSON and markdown timelines.

The aggregation keeps model_call_count truthful: fallback-only runs still report zero model-call rows, and cycle totals skip fallback actuals for linked evals that already have model-call rows to avoid double counting.

Verification

• cargo test -p xvision-engine eval::report::tests::aggregate_run_token_totals_falls_back_to_eval_run_actuals
• cargo test -p xvision-engine autooptimizer::cycle_runs::tests
• cargo test -p xvision-engine autooptimizer::cycle_export::tests
• cargo test -p xvision-cli commands::eval::tests::status_line_fallback_tokens_render_without_model_calls
• git diff --check

Review

• Correctness review: no findings after fixes.