Stale per-repo issue-discovery fields survive a successful scoring round and inflate roll-up totals

[kevin8600]

Description

run_issue_discovery has two per-round reset paths for a miner's issue-discovery data, and they diverge:

• The no-issues path (_clear_issue_discovery_fields, scan.py) zeroes the issue fields on every repo_evaluation.
• The scoring path (_finalize_repo_issue_scores) only rewrites the repos seen this round (sorted(set(repo_acc) | set(open_counts))), then _roll_up_issue_totals sums over all repo_evaluations.

As a result, a repo that was issue-scored in a prior round but is absent this round keeps its stale per-repo issue values, and those get summed back into the round-level totals.

This is reachable because an OSS-fetch-failed miner is restored from the evaluation cache (store_or_use_cached_evaluation in neurons/validator.py replaces the eval with the cached one, which carries prior-round repo_evaluations). Issue discovery then runs on that restored eval — and forward.py explicitly expects it to produce fresh issue-discovery fields — but the scoring path leaves the now-inactive repo stale.

Steps to Reproduce

1. Round N-1: a miner (with ≥1 merged PR, so it's cacheable) is issue-scored in repo A → the cache stores repo_evaluations['A'] with non-zero total_solved_issues / issue_discovery_score / is_issue_eligible.
2. Round N: the miner's OSS PR fetch fails (github_pr_fetch_failed), so the evaluation is restored from cache (carrying repo A's stale issue fields).
3. Round N: the issue-mirror fetch succeeds, but the miner now only has issues in a different repo B (and no open issues in A), so A is absent from both repo_acc and open_counts.
4. _finalize_repo_issue_scores rewrites only B; repo A is never touched; _roll_up_issue_totals sums A (stale) + B (fresh).

Expected Behavior

The scoring path zeroes stale per-repo issue fields for repos not seen this round (same as the no-issues path), so the round-level roll-up reflects only this round's issues.

Actual Behavior

Repo A retains total_solved_issues=7, issue_discovery_score=8.12, is_issue_eligible=True; the round-level total_solved_issues / issue_discovery_score / is_issue_eligible inherit those stale values. The corrupted per-repo rows are persisted via bulk_store_evaluation and re-written to the evaluation cache via update_issue_discovery, so the error compounds across rounds. (Current-round emission allocation reads the freshly-rebuilt issue_discovery_issues list, so this is a stored-data / reporting correctness bug — CLI miner score, analytics, and the cross-round cache — rather than an emissions bug.)

Environment

• OS: Linux (CI) / cross-platform
• Python version: 3.12
• Commit/Version: 822056d (v5.0.0)

Additional Context

The no-issues path already does the correct full reset (_clear_issue_discovery_fields loops over every repo_eval), which shows the intended contract; the scoring path simply drifted from it. Fix: extract the per-repo reset into a shared helper and apply it on both paths.