XTester v.0.0.70

Reliable autonomous strategy research, convenient version navigation, and a new Git diff preview.

AI Research Lab:

• The research loop no longer continues known-dead branches: the base version is chosen by
  hard-requirement compliance, runtime telemetry quality and objective progress, and finishing
  a research run requires a fresh Walk-Forward / Out-of-Sample check.
• Added semantic checks of the generated C# before backtesting: unreachable entries, look-ahead
  in Donchian breakouts, wrong funnel counter order, inconsistent ATR units, double risk scaling,
  and a missing protective exit.
• Plateau mode now truly waits for the user's choice. Auto-continue is allowed only after the
  regular iteration limit; questions and ordinary chat messages are not lost, and direct
  instructions feed into the next hypothesis.
• Restored the cumulative LLM token counter after reopening a research; legacy data is normalized
  without resetting. Regular [HALT] risk-stop events are no longer counted as exceptions.
• Fixed the false "m15 data unavailable" message after warm-up and an attempt to parse
  a PlainText/C# model response as JSON.

Lab navigation & UI:

• "All / Working / Broken" filters above the version tree. Filtering keeps the ancestor path
  to a matching strategy while unrelated dead branches are hidden.
• Added a sortable table view with summary metrics, Fitness delta and genome parameter search;
  switching between the tree and the table adapts the panel width without overlapping buttons.
• Fixed the Director chat message queue, selected version persistence, pause state restore,
  and premature skipping of the interactive plateau choice.

Git & windows:

• Git history got a colored line-by-line diff preview (WebView2) with correct line numbers
  and a stable changed-files table.
• Source control keeps a single primary Commit button; "Commit & push" and "Commit to a new
  branch" moved into a compact menu.
• Improved window geometry restore across monitor and DPI changes.

Reliability:

• The lab's headless backtest reliably restores the shared simulator lifecycle;
  step-size/min-notional rejections and real fills are correctly reflected in diagnostics
  and candidate selection.
• Extended the regression suite for AI Lab, Git diff, localization and window geometry.