perf(validator): free per-PR file content after scoring

[anderdc]

Summary

score_pr attaches each PR's full source text to its ScoredPR via scored.files — MirrorFile.head_content / base_content, up to ~1 MB per file. That content is only needed transiently to compute the tree-diff scalar scores (token/structural/leaf counts + base score), but it currently stays attached to every ScoredPR and is held across all miners in miner_evaluations for the entire scoring round.

This frees the heavy text as soon as the scalar scores are extracted, keeping only the lightweight file metadata (filename / additions / deletions). The persistent cache already does exactly this — _scored_mirror_pr_for_cache sets files = None (classes.py); this just applies the same treatment to the live round dict.

Effect: peak scoring-round memory drops from roughly miners × PRs × files × content down to metadata + a single PR's content at a time, with no change to scoring output.

Why it's safe

The file content is consumed entirely within score_pr (the file_contents dict built by mirror_files_to_legacy and passed to calculate_base_score_for_pr_files). Nothing downstream reads head_content / base_content after that point:

• _calculate_pr_multipliers reads only PR metadata + scalar scores
• finalize_miner_scores operates on scalar scores
• bulk_store_evaluation → get_all_file_changes reads only file metadata (discards content)

Type of Change

• Performance / optimization

Testing

• Existing mirror scoring tests pass (72 passed)