[irvineoy]

Replace the additive compilation/correctness/speedup score with max(0, 1 - 1.05 / speedup_ratio) ** 0.5. Correct, comparable candidates earn a positive score only above 1.05x speedup; failed checks and invalid comparisons score zero. The existing score field and mean-per-case speedup aggregation remain in use.

Document SCORE_POWER and SCORE_SPEEDUP_THRESHOLD, curve examples, and historical score comparability. Text reports and the dashboard display up to four decimal places. For an illustrative 1.25x speedup, the old score of 245 becomes 0.4; these are formula examples, not new GPU measurements.

The branch is rebased onto current main and adds two follow-up commits:

• Tests added on main after this branch was cut (tests/test_task_run_v2.py, tests/test_candidate_verification.py) still asserted the 120/220/320 scale; they now assert the bounded score. The review-gate fixture uses a synthetic 2x candidate, so its score assertion still distinguishes accepted from rejected candidates instead of reading 0 for both.
• Score cases that repeated an already covered branch are removed: equivalent invalid speedups and constants, a default-curve point duplicated by the power test, method-consistency inputs that have dedicated tests, and the all-zero run of the report-format integration test.
• The per-measurement average score in text reports uses four decimals, matching the other score fields.
• src/utils/utils.py and src/utils/report_generation.py are removed. They had no callers, imports, or documentation references, and they inferred compilation and correctness from score thresholds (20/120/220) that the bounded score no longer encodes. Current reports count those from pass_compilation and pass_correctness.

Validation (CPU only):

• Linux x86_64, Docker python:3.12-bookworm with Python 3.12.15 and CPU PyTorch 2.14.1, using the CI dependency set and commands on this branch (80dca1e) and on main (c0da09d) from the same checkout bundle:
  • make check-perf-helpers: pass on both.
  • python -m pytest -q tests --durations=25: this branch 14844 passed, 143 skipped, 0 failed; main 14819 passed, 143 skipped, 0 failed.
  • python -m compileall -q src tasks tests: pass on both.
• macOS arm64, Python 3.12: the focused scoring, evaluator, held-out, task lifecycle, candidate verification, reporting, and visualization tests passed (156). The full suite showed no failure on this branch that does not also occur on main (Linux-only process containment and GNU tool assumptions); three timing-sensitive process tests failed only while both suites ran concurrently and passed when rerun alone.
• No GPU evaluation was run. This change re-scores evaluator outputs and does not change measurement, so no before/after task evaluation is included; scores from earlier runs are not directly comparable with new ones.