fix: prevent scoring data loss and fix fragile test in stale PR refresh

[wdeveloper16]

Description

Two gaps left unaddressed by #769.

Gap 1 — Stale UPSERT silently zeroes previously-scored rows

store_pull_requests_bulk in gittensor/validator/utils/storage.py:74-76 routes stale PRs through BULK_UPSERT_PULL_REQUESTS, which uses ON CONFLICT DO UPDATE SET with EXCLUDED.* — overwriting every column on conflict. PullRequest.from_graphql_response() leaves all scoring fields at dataclass defaults (base_score = 0.0, earned_score = 0.0, credibility_multiplier = 1.0, etc.). Any PR that was previously scored — for example, one that was OPEN with collateral before the window shifted — has those computed values silently replaced with zeros on every subsequent scan.

Gap 2 — Fragile positional assertion in test_stale_closed_prs_are_stored_separately

tests/validator/utils/test_storage_mirror.py:176 asserts stale storage correctness via call_args_list[3]. If the call order inside store_evaluation ever changes, the assertion silently targets the wrong call and the test continues to pass.

Changes

• Add REFRESH_STALE_PR_STATES to queries.py — a targeted UPDATE that touches only pr_state and updated_at, leaving all scoring columns on existing rows untouched
• Add Repository.refresh_stale_pr_states() backed by the new query
• Wire store_evaluation to call refresh_stale_pr_states instead of store_pull_requests_bulk for the stale bucket
• Replace call_args_list[3] with refresh_stale_pr_states.assert_called_once() and add a leak guard asserting stale PRs do not reach store_pull_requests_bulk

No schema changes required.

Closes #991

[wdeveloper16]

Hi, @anderdc
Can you review the PR, please?

[wdeveloper16]

• store_evaluation routes stale_closed_pull_requests through store_pull_requests_bulk, which runs BULK_UPSERT_PULL_REQUESTS — an ON CONFLICT DO UPDATE SET that overwrites every column including earned_score, base_score, credibility_multiplier, pioneer_rank, and pioneer_dividend.
• PullRequest.from_graphql_response() never computes scoring fields for stale PRs, so those columns arrive as dataclass defaults (0.0 / 1.0).
• The overwrite is deterministic: any PR scored in epoch N becomes stale in epoch N+35 as the lookback window shifts, and the next scan silently zeroes its stored scores — destroying audit-accurate scoring history every cycle for any miner with aged-out PRs.

This PR replaces the full UPSERT for the stale bucket with a dedicated REFRESH_STALE_PR_STATES query that touches only pr_state and updated_at, leaving all scoring columns untouched. No schema changes, no new tables, no behavioural change to the scoring pipeline.