Fix codeinsights ast_distance / asm_distance returning similarity not distance

[Kymi808]

Summary

codeinsights_code_evaluation_metrics.py imports

from Levenshtein import ratio as levenshtein_distance_ratio

then in ASTAnalyzer.calculate_ast_distance (and the assembly path in CodeInsightsCodeEvaluationMetric.evaluate_generation) returns it directly:

return levenshtein_distance_ratio(nodes1, nodes2)   # AST
asm_distance = levenshtein_distance_ratio(gen_asm, truth_asm)

Levenshtein.ratio is a similarity in [0, 1] (1.0 = identical, 0.0 = completely different), but everything around the call site assumes distance semantics:

signal	says
function name calculate_ast_distance	distance
docstring	"a float in [0,1], where 0 means identical ASTs and 1 means completely different (or a parse failure)"
stat names MetricName("ast_distance") / "asm_distance" (contrasted with the sibling "codebert_similarity" in the same class)	distance
error fallbacks (return 1.0 for parse failure, asm_distance = 1.0 for compile failure, comment "max distance for errors")	distance

So a perfect AST or ASM match was reported as 1.0, indistinguishable from a parse / compile failure, and the ordering was inverted (higher score looked better than lower).

Fix

Convert similarity to distance at the two call sites:

return 1.0 - levenshtein_distance_ratio(nodes1, nodes2)
asm_distance = 1.0 - levenshtein_distance_ratio(gen_asm, truth_asm)

The legacy batch helper evaluate_ast_distances_batch calls analyzer.calculate_ast_distance(...) and so picks up the fix transitively.

Why this isn't a behavior change someone would be relying on

• No other file in the repo references ast_distance / asm_distance / normalized_distance — no presentation/schema, no higher_is_better/lower_is_better annotation, no test.
• The error fallbacks already use 1.0 as "max distance", so any consumer that read those error values was already assuming distance semantics.

Test plan

Added src/helm/benchmark/metrics/test_codeinsights_code_evaluation_metrics.py. Uses pytest.importorskip for the optional codeinsights extras (clang.cindex, Levenshtein) and stubs clang.cindex.Index.create + _extract_ast_features so the tests don't need a real libclang install:

• identical inputs → 0.0 (matches docstring)
• totally different inputs → 1.0
• partial overlap → small positive in (0, 1)

The first two fail on main (return 1.0 and 0.0 respectively) and pass with this change. pytest -o addopts="" on the new file → 3 passed; black --check clean on both files.