fix(detection): resolve Kubernetes compare regressions

[mstykow]

Summary

• fix the real regressions exposed by the kubernetes/kubernetes compare-outputs --profile common run, including embedded-PEM source scanning, local license-reference cleanup, Conda false positives on generic *env*.yaml, and holder extraction for plain year-range copyright lines
• add focused regression coverage for the Kubernetes-driven fixes, rerun the affected copyright golden, and keep the remaining Provenant-better or non-regression deltas documented as accepted outcomes
• add docs/BENCHMARKS.md as the canonical package-detection verification record and simplify the package-detection scorecard/docs to point detailed compare results there

Scope and exclusions

• Included:
  • scanner text-detection fixes for source files that embed PEM blocks
  • post-processing cleanup for local LICENSE references that should not retain unknown-license-reference noise
  • Conda environment parsing guardrails to avoid package false positives on generic env-style YAML fixtures
  • copyright holder extraction improvements for plain year-range lines
  • package-detection benchmark documentation and scorecard simplification
• Explicit exclusions:
  • deleting docs/implementation-plans/package-detection/PARSER_VERIFICATION_SCORECARD.md
  • adding new benchmark targets beyond the currently verified Boost and Kubernetes examples

Intentional differences from Python

• keep Provenant's broader Dockerfile and go.work package coverage where it is more correct than ScanCode
• keep cleaner license-expression, URL, and metadata outcomes where the remaining differences are accepted non-regressions rather than parity losses

Expected-output fixture changes

• Files changed: testdata/copyright-golden/copyrights/misco2/ocaml.txt.yml
• Why the new expected output is correct: the improved year-range holder extraction now keeps the holder attached to Copyright 2013-2020 by OCamlPro, so the golden should record both the fuller copyright line and the OCamlPro holder instead of underreporting the result