fix: repair broken error handling, deprecated model, and Windows crashes

[Pyhroff]

What changed

• TAP and Crescendo now catch exceptions around API calls instead of crashing the whole run on a transient failure (rate limit, timeout, etc.) — matches PAIR's existing error handling.
• Fixed ci/sweep silently folding errored goals into the ASR denominator, which could misreport a failed scan as 0% ASR ("model resisted") instead of flagging that the scan never actually ran. ci now fails the gate on any errored goal, and sweep shows an explicit error count instead of a false RESILIENT verdict.
• Replaced the deprecated/retired Groq model id (llama-3.3-70b-versatile) with a currently valid one everywhere it was hardcoded (.env.example, config.py, CLI option defaults, docstrings).
• Forced UTF-8 stdout/stderr on Windows so rich's box-drawing/arrow characters don't crash on the default cp1252 console codepage.
• Fixed defend's instruction_override regex missing stacked modifiers (e.g. "ignore all previous instructions" wasn't matched — only single-modifier phrasing was).
• Fixed history crashing on sqlite3.Row.get() (Row doesn't support .get(), only dict does).
• Fixed TAP's scan-result formatting — garbled 24n d3 now reads 24 nodes (d3).
• Added a category-filter error message listing valid categories instead of a bare "no behaviors matched."
• Added a regression test for the errored-goals-must-fail-the-ci-gate behavior.

Why

Ran a full manual audit of every CLI command against the live Groq API. Several of these were only exposed once real API calls were actually made (a rate-limit error during a live Crescendo run crashed it uncaught, which is what led to auditing and fixing TAP's identical gap before it could do the same). The ASR/errored-goals issue is the most consequential fix — as shipped, a ci gate could silently PASS when the scan never actually completed, which defeats the point of a CI safety gate.

Testing

• Full test suite: 62/62 passing (61 existing + 1 new regression test)
• Every CLI command (scan with pair/tap/crescendo, sweep, ci, defend, history, report, serve, gcg) exercised against the live Groq API and/or a local HF model, not just imported
• Confirmed the exact bug scenarios before and after each fix (e.g. reproduced the regex false-negative, the sqlite3.Row crash, and the ERROR-displayed-as-RESILIENT sweep output, then verified each was resolved)