fix(issue-discovery): solving-PR cache poisoned by unscored OSS PRs (token_score=0 blocks valid solves) #1677

Description

_build_solving_pr_cache in gittensor/validator/issue_discovery/scan.py pre-populates the cross-miner solving-PR cache from every entry in evaluation.merged_prs, including PRs whose OSS score_pr() call early-returned without ever tokenizing.

Those incomplete PRs keep the ScoredPR defaults (token_score=0, base_score=0, files=None). Once cached, issue discovery treats them as a cache hit and never calls MirrorClient.get_pr_files — so a solving PR that would score on a fresh files fetch stays at token_score=0 and fails min_token_score_for_valid_issue.

This is the OSS → cache sibling of closed #836 (which fixed fabricated zeros on the cache-miss path inside _resolve_solving_pr_score). #836 correctly refuses to cache scoring_data_stored=False on fetch; the pre-populate path still injects the same class of fabricated zero from unscored merged_prs.

Distinct from open #1616 (unique_repos_contributed_to undercount on the same early-return paths) — that bug is about repo attribution; this one is about issue-discovery score resolution via the solving-PR cache.

Steps to Reproduce

1. Construct a MinerEvaluation whose merged_prs contains a MERGED ScoredPR that never completed OSS scoring, e.g. scoring_data_stored=False caused score_pr to return early:

   • scored.files is None
   • scored.token_score == 0.0
   • scored.base_score == 0.0
   • PR remains in merged_prs (eligibility already passed at load time)

2. Call _build_solving_pr_cache({uid: evaluation}).

3. Observe cache[(repo, pr_number)] == CachedSolvingPR(base_score=0.0, token_score=0.0).

4. Run issue discovery for a miner with a solved issue whose solving_pr.pr_number is that PR. _resolve_solving_pr_score hits the cache and does not call get_pr_files.

5. Even if the mirror would now return scoring_data_stored=True with real file contents, the issue is treated as below-threshold / unscored.

Representative flow on current test:

OSS score_pr (mirror/scoring.py):
  scoring_data_stored=False and no fixed_base_score → return early
  ScoredPR stays at defaults; PR remains in merged_prs

_build_solving_pr_cache (scan.py):
  caches every merged_prs entry, including the incomplete one
  → cache[(repo, N)] = CachedSolvingPR(0.0, 0.0)

_resolve_solving_pr_score:
  key in cache → HIT → return zeros → never fetch files

valid-solved gate:
  token_score (0) < min_token_score_for_valid_issue → no discovery score

Contrast with the intentional design pinned by test_all_merged_prs_cached_regardless_of_token_score: caching a genuinely tokenized token_score=0 PR is correct (saturation / low-token suppression). Caching an unscored PR that merely looks like zero is not.

Expected Behavior

Only pre-populate the solving-PR cache from PRs whose OSS scoring actually completed, for example:

• scored.files is a non-empty list (tokenization ran), or
• scored.base_score > 0 (e.g. fixed_base_score path completed)

Incomplete early-returns (files is None / empty fetch with no fixed base, base_score still 0) must remain cache misses so _resolve_solving_pr_score can fetch files (and still honor #836: do not cache scoring_data_stored=False).

Actual Behavior

Every merged_prs entry is cached, including unscored defaults. Cache hits suppress get_pr_files, under-crediting issue discovery whenever mirror file data was pending/flaky during the OSS pass.

Environment

• OS: macOS
• Python version: 3.12.13
• Commit/Version: test @ 26ebeb8

Additional Context

Affected code:

• gittensor/validator/oss_contributions/mirror/scoring.py — score_pr early returns (~139–140, ~147–149, ~174–176)
• gittensor/validator/issue_discovery/scan.py — _build_solving_pr_cache (~397–416), _resolve_solving_pr_score (~618–641)
• tests/validator/issue_discovery/test_scan.py — test_all_merged_prs_cached_regardless_of_token_score (documents intentional zero caching for scored PRs; needs a sibling test for unscored PRs)

Suggested regression test:

1. Seed merged_prs with an incomplete ScoredPR (files=None, scores 0).
2. Build cache → key must be absent.
3. Drive _resolve_solving_pr_score with a mock get_pr_files that returns real files → must be a miss, fetch, and populate real scores.

I am planning to open a PR targeting test for this fix.