Improve self-consistency answer extraction and voting

[lukeinglis]

Summary

• Safe default projection: Default now uses strip().lower() instead of just .strip(), adding case normalization while keeping behavior predictable. Smart answer extraction is available via opt-in presets.
• Built-in presets: Add ProjectionPreset StrEnum (MATH, GENERAL, EXACT) and projection_preset parameter to SelfConsistency.__init__() so callers can opt into smart answer extraction without writing custom functions.
  • MATH: Extracts \boxed{} content with nested brace handling
  • GENERAL: Tries boxed extraction, explicit answer patterns ("Final Answer:", "Therefore..."), last paragraph fallback — per Wang et al., 2022
  • EXACT: Legacy .strip() behavior (no lowercasing)
• Fix single-pattern regex: create_regex_projection_function now returns a plain string (not a 1-tuple like ('45',)) when given a single pattern, producing clean response_counts keys
• Shared _extract_boxed helper: Consistent nested-brace handling used by both the MATH preset and the _extract_answer function, eliminating behavioral divergence
• Tests: 15 new tests covering presets, _extract_boxed, _extract_answer, Enum validation, and custom function priority

Motivation

Per Wang et al., 2022, self-consistency works by sampling diverse reasoning chains, extracting the final answer from each, and taking a majority vote over the extracted answers. The previous default .strip() projection means every caller must supply their own projection function to get correct voting behavior.

Rather than making the default aggressively heuristic (which would be a breaking change), this PR keeps the default simple (strip().lower()) and provides opt-in presets for smart extraction. Users who want paper-correct behavior can use projection_preset="general" or projection_preset="math".

Additionally, create_regex_projection_function returned 1-tuples for single patterns (e.g. ('45',) instead of '45'), which created awkward keys in response_counts that every consumer had to unwrap.

Backwards compatibility

• Callers passing a custom consistency_space_projection_func are unaffected — explicit functions still take priority over presets
• The default behavior change is minimal: strip() → strip().lower(). Only case-sensitive comparisons would see different results.
• Callers needing exact legacy behavior can use projection_preset="exact"
• Smart answer extraction is opt-in via projection_preset="general" or projection_preset="math"

Test plan

• Boxed answer extraction (\boxed{42} → "42", nested braces \boxed{x^{2}+1} → "x^{2}+1")
• Explicit answer pattern extraction ("Final Answer: 45" → "45")
• Case normalization ("YES" → "yes")
• Single regex returns string, not tuple
• Multi regex still returns tuple for hierarchical voting
• ProjectionPreset is a proper StrEnum
• All presets initialize correctly (math, general, exact)
• "exact" preset preserves legacy .strip() behavior
• Invalid preset raises ValueError
• Custom function takes priority over preset
• _extract_boxed returns None when no boxed expression found
• _extract_answer falls back through patterns correctly
• Ruff linting passes
• All 42 self-consistency tests pass

[beatsmonster]

@lukeinglis please rebase to main/v1