Fix dogfood proof discovery at scale

[rcs2153]

Summary

Fix the repo-local governed phase runner so phase-close reads approval-presentation records from the target run's canonical SHA-256-partitioned directory instead of scanning every run in the shared store.

Root cause

The helper scanned the complete approval-presentation record tree and failed closed after 250 records. Once normal dogfood history exceeded that global bound, a valid target-run presentation proof became undiscoverable even though Core already stores records under a per-run hash directory.

Changes

• derive the same SHA-256 run partition used by the local state backend;
• read only the requested run bucket;
• retain a 250-record fail-closed ceiling within that bucket;
• add regressions for more than 250 unrelated records and an oversized target bucket.

Validation

• npm run test:dogfood-helper (32 passed)
• npm run check:docs
• npm run check:integrations
• git diff --check
• real dg/blocker phase-close returned approval_presentation_enforcement: proof_enforced

Governed phase

• implementation workflow: dg/blocker
• run: run-1786536294355921000-2
• approval: approval/run-1786536294355921000-2/fix-approved
• presentation proof: presentation/7556ba28af9f5510
• outcome: granted; 39 events; 0 retries; 0 escalations
• PR workflow: dg/pr
• PR run: run-1786537818969755000-2
• PR approval: approval/run-1786537818969755000-2/pr-scope-approved

Boundary

No Core runtime or approval semantics, persistence format, workflow specs, schemas, provider behavior, report artifacts, hosted behavior, external writes, or release posture changed. Repository edits, checks, git operations, and this PR were performed outside the kernel and are disclosed here.