Fix #84: enforce 4/6-message conversation depth by fidelity

[RandomOscillations]

Summary

• fix conversation execution to use message budgets (medium=4, high=6) instead of truncating at 2/3 messages
• mark only the final message as final, preserving proper back-and-forth structure
• add execution tests with mocked LLM calls to assert message count, speaker alternation, and final-flag behavior

Validation

• pytest -q tests/test_conversations.py

Closes #84

[DeveshParagiri]

Code Review

Verdict: ✅ Ready to merge

Summary

This PR correctly fixes the conversation depth bug by iterating over the correct number of messages (4 for medium, 6 for high fidelity) instead of the mislabeled turn count (2/3).

Verification Checklist

Requirement	Status
Medium fidelity = 4 messages	✅
High fidelity = 6 messages	✅
Low fidelity = 2 messages	✅ (defensive fallback)
Only final message marked is_final=True	✅
Proper speaker alternation	✅
Tests validate behavior	✅

Key Fixes

1. Changed loop from range(turns) where turns=2/3 to range(total_messages) where total_messages=4/6/2
2. Simplified is_final logic to only mark the true final message
3. Clean helper function _message_count_for_fidelity()

Tests

Both new tests pass and validate message count, speaker alternation, and final flag behavior.

No changes required.