red team: _MAX_BACKTRACKS = 10 is hardcoded; should scale with total_turns

[Aryansharma28]

Problem

`python/scenario/red_team_agent.py:209`:

```python
self._MAX_BACKTRACKS = 10
```

This is unconditional. Edge cases:

• `total_turns=5` → 10 backtracks possible but only 5 turns exist → cap is meaningless
• `total_turns=100` → 10 backtracks against 100 turns means we waste 90% of the budget on dead ends if the target is highly defensive

A backtrack also costs a turn from the budget (documented in `marathon_script` docstring), so they have direct cost.

Why it matters

Real targets vary wildly in defensiveness. A well-aligned model may force 20+ backtracks in a 50-turn run. A weakly-aligned model may need zero. Hardcoding 10 splits the baby badly for both extremes.

Proposed fix

Either:

• `max(1, total_turns // 3)` as the implicit default, OR
• Expose as a constructor parameter `max_backtracks: Optional[int] = None` defaulting to the formula above.

Same in TS `backtracksRemaining` initialization.

Acceptance

• Default scales with `total_turns`
• Optional explicit override
• Test: `total_turns=3` doesn't allow more backtracks than turns

Discovered during PR #306 review.

[Aryansharma28]

Tracked under EPIC: langwatch/langwatch#1713 (Scenarios Red Teaming)

[Aryansharma28]

Implemented in #347 (merged into #346).

• max_backtracks parameter added to RedTeamAgent.__init__ (Python) / RedTeamAgentConfig (TS).
• Default auto-scales as max(1, total_turns // 3) — 5-turn runs get 1, 30-turn runs get 10 (matches the old hardcode), 100-turn runs get 33.
• Negative values raise; explicit override works.
• Acceptance criteria met: default scales with total_turns, optional explicit override, test coverage verifies total_turns=1 still gets at least 1 backtrack.