Load master repo weights from on-chain registry

[LandynDev]

Part of repo-registration phase 2a (task 2a-5). Stacked on #1667 — merge order #1665 -> #1666 -> #1667 -> this.

• New repo_registry/loader.py: RegistryLoader reads the registry once per scoring round at the snapshot block hash (atomic across validators), decodes the FP6 hyperparam table (keys 1-26), memoizes per snapshot, and maintains a last-good disk cache
• Identity shim at the load_weights.py boundary: contract github_id resolves to lowercase owner/name dict keys — downstream scoring sees zero changes; renames appear as a new key next snapshot (forward-only, no history migration)
• Fallback ladder (spec semantics): contract @ snapshot -> last-good disk cache -> baked JSON. Paused/empty registry deliberately skips the cache and lands on baked so a set_paused rollback is not defeated by stale cache. Worst case = today's behavior.
• Proof enforcement is structural: verified no code path filters or renormalizes on App-install status; share vectors derive only from contract state at the snapshot hash (vtrust byte-identity)
• issue_discovery_share default fix: DEFAULT_ISSUE_DISCOVERY_SHARE 0.5 -> 0.0. Bug: novel repos entering via consensus inherited the 0.5 default, silently diverting half their slice to issue discovery (17/18 live repos use 0.0). All baked entries set the field explicitly, so baked behavior is unchanged. Regression tests included. (Checked das-gittensor: only aggregates issue_discovery_score DB columns, no mirrored default constant — no cross-repo change needed.)
• Out-of-bounds contract params raise on parse and degrade down the ladder instead of crashing — defense in depth over the contract's own bounds

Tests: +26 (ladder legs, shim + rename, snapshot atomicity, discovery-share regression, novel-repo consensus entry); full suite 1123 passed; ruff clean.