FIX: separator normalization for harm categories and likert invariants

[Saurabh Yadav (onlyysaurabh)]

Fast-follow to PR #2663 addressing the residual items for Issue #2661 highlighted by varunj-msft:

1. Separator Normalization: PR FIX: strict scorer-eval category match + CSV-axis test #2663's casefold() fast path handled case-only differences (e.g., Jailbreak vs jailbreak), but separator variations (HateSpeech vs hate_speech, ProtectedMaterial vs protected_material) still failed matching without hand-added aliases.
2. Parametrized Likert Invariant: Shipped Likert presets were only partially tested (2 of 8 presets asserted against hardcoded string literals).
3. Metrics Invariance Guard: Post-evaluation assertions on metrics.harm_category were missing, allowing silent None regressions on direct-call paths.

Summary of Changes

1. Canonical Taxonomy Separator Normalization (pyrit/models/harm_category.py):
   • Populated stripped variants (key.translate(_SEP_TRANS) where _SEP_TRANS = str.maketrans("", "", "_- ")) into _CANONICAL_LOOKUP alongside exact names and display values.
   • HarmCategory._resolve_canonical_category performs exact lookup first, then falls back to stripped lookup.
   • Resolves CamelCase variants (HateSpeech -> HATE_SPEECH, ProtectedMaterial -> PROTECTED_MATERIAL, SelfHarm -> SELF_HARM) automatically without needing manual entries in _HARM_CATEGORY_ALIASES.
2. Evaluator Fast-Path Separator Normalization (pyrit/score/scorer_evaluation/scorer_evaluator.py):
   • Strips _, -, and spaces from both sides in _score_matches_harm_category fast path: c.casefold().translate(_SEP_TRANS) == target_clean.
   • Enables unrecognized/custom taxonomy categories (PromptInjection vs prompt_injection, jailbreak-attempt vs jailbreak_attempt) to match cleanly.
3. Parametrized Likert Invariant (tests/unit/score/test_self_ask_likert.py):
   • Replaced the 2-preset test with a parametrized invariant across all 8 presets in LikertScalePaths that have evaluation_files is not None.
   • Asserts scale.category == preset.evaluation_files.harm_category.
4. Metrics Invariance & Rejection Tests (tests/unit/score/test_scorer_evaluator.py):
   • Added assertions on metrics.harm_category (assert metrics.harm_category == "hate_speech" and assert metrics.harm_category == "violence") in test_run_evaluation_async_* to guard against None regressions.
   • Added positive tests covering CamelCase, kebab-case, snake_case, and spaced formats.
   • Added negative tests verifying that prefix/substring traps (harm vs self_harm, self vs self_harm, speech vs hate_speech, code vs insecure_code, cyber vs cyberattack) are strictly rejected with ValueError.
5. Taxonomy Unit Tests (tests/unit/models/test_harm_category.py):
   • Added parametrized tests asserting HarmCategory.parse cleanly maps CamelCase, kebab-case, snake_case, and space variants directly into enum members.

Tests and Documentation

• Targeted Unit Tests:
  uv run pytest tests/unit/score/test_scorer_evaluator.py tests/unit/score/test_self_ask_likert.py tests/unit/models/test_harm_category.py -q
  -> 171 passed in 6.00s
• Subsystem Test Suite:
  uv run pytest tests/unit/score/ tests/unit/models/ -q
  -> 3,201 passed, 0 failures, 0 regressions
• Code Quality:
  uv run ruff check -> Passed (0 warnings, 0 errors)
  uv run ruff format --check -> Passed (All 5 files formatted)