fix: hybrid-merge cached mirror PRs when only mirror fetch fails

[mkdev5]

Summary

Mixed legacy + mirror miners are silently zeroed when mirror.gittensor.io is down. combine OR's mirror_eval.fetch_failed into the legacy github_pr_fetch_failed, but should_use_cache_fallback still requires total_prs == 0 — so a successful legacy fetch (1+ PR) blocks the fallback and the validator finalizes a partial evaluation missing every mirror PR. Reach: any miner with PRs in both an entrius/* mirror repo and any non-mirror repo, during a mirror outage. Drops past MIN_VALID_MERGED_PRS → 0 OSS reward.

Adds a per-source mirror_pr_fetch_failed flag (mirror is binary — atomic load), wires it through combine, and routes mixed mirror-only failures through a hybrid merge in store_or_use_cached_evaluation: keep the fresh legacy data, splice cached mirror PRs back in, and re-cache the result. Cached ScoredMirrorPR reuse is safe because base_score/token_score are deterministic over the merged diff; credibility and other cross-PR multipliers get recomputed in finalize_miner_scores.

The pre-mirror legacy partial-pagination semantic (#516) is preserved — that path uses the existing should_use_cache_fallback gate untouched. Pure-mirror miners and full-failure cases also keep the existing full-eval fallback.

Related Issues

Closes #681

Type of Change

• Bug fix
• New feature
• Refactor
• Documentation
• Other (describe below)

Testing

• Tests added/updated
• Manually tested

Checklist

• Code follows project style guidelines
• Self-review completed
• Changes are documented (if applicable)

[anderdc]

The bug is real and the diagnosis is correct, but the hybrid-merge approach is more invasive than the situation warrants given the legacy path is on the chopping block.

The whole combine module is documented as going away on delete-day ("On delete-day this whole module goes away — at that point MirrorMinerEvaluation becomes the canonical container"). Once legacy is gone:

• mirror_pr_fetch_failed collapses back into a single failure flag
• The mixed-source problem stops existing entirely
• A mirror outage just means "no data this round, use cached eval" — same shape as legacy works today

Everything the hybrid splice adds (_splice_cached_mirror_into, the dedicated conditional block, the re-cache step) is transition-only scaffolding that we'll then have to delete. And during the transition it produces a mixed-snapshot eval (fresh legacy PRs + stale-cached mirror PRs evaluated together) where credibility and other cross-PR multipliers get recomputed across a Frankenstein view of the miner's state.

Suggested simpler fix:

1. Keep the mirror_pr_fetch_failed flag and the combine wiring — that part is necessary signal.

2. In store_or_use_cached_evaluation, just let mirror failures through the existing cache-fallback path instead of building a new branch. Roughly:

   if not miner_eval.should_use_cache_fallback and not miner_eval.mirror_pr_fetch_failed:
       bt.logging.warning(...)
       continue
   
   cached_eval = self.evaluation_cache.get(...)
   if cached_eval is not None:
       bt.logging.info(...)
       miner_evaluations[uid] = cached_eval
       cached_uids.add(uid)
   

3. Drop _splice_cached_mirror_into and the dedicated hybrid block in neurons/validator.py entirely.

The trade-off: during a mirror outage, mixed-source miners get evaluated against a coherent one-round-stale snapshot instead of a fresh-legacy + stale-mirror hybrid. They lose the freshness of this round's legacy PRs for one cycle. Given mirror outages are transient and we're heading toward mirror-only anyway, that's the right trade — less code, no cross-source contamination, and nothing extra to delete on delete-day.

Tests: test_combine.py assertions stay; the validator-side tests can be simplified to assert "mirror failure with cached eval available → swapped to cached eval, returned in cached_uids."

[mkdev5]

pushed the update