fix: recover failed repeated semantic source drift safely

[acidkill]

Summary

Allow explicit, audited stale-source recovery when a semantic discovery checkpoint fails again after an earlier recovery. This unblocks the production run with 19/20 strategies complete without silently discarding progress.

Changes

• Accept failed only with the existing version-3 source manifest, fresh source-change proof, zero semantic graph effects, exact prior audit history, lease and CAS fences.
• Atomically change failed to paused while resetting only unapplied discovery; preserve original reference time and other strategy counters.
• Support a repeated prior stale-source marker and exact committed retries.
• Add negative unit cases and a SurrealDB 3.2.4 CAS integration regression.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• Tests

Testing

• Unit tests added/updated
• Integration tests added/updated
• Manual production recovery: pending independent review and CI

PYTHONPATH=src SURREALDB_URL=ws://127.0.0.1:18082/rpc SURREALDB_USER=root SURREALDB_PASS=root pytest -q tests/integration/test_surrealdb_semantic_recovery_cas.py tests/unit/test_stale_semantic_recovery_repeat.py tests/unit/test_surrealdb_consolidation_state.py: 20 passed against disposable local SurrealDB 3.2.4. Ruff and mypy passed on the implementation commit; CI is authoritative for the final head.

Checklist

• Code follows project style
• Tests prove the fix
• Independent PR review and green CI before merge
• Documentation/changelog: no public API or release change; internal exceptional recovery path only

Related Issues

Production run 770a1a98-1483-419c-a72c-9a21cfce7c74 is failed in semantic_link_discovery_neurons; no graph-effect counters were applied.