fix(cluster): require recorded history evidence

[FreeOnePlus]

Summary

• require at least two recorded time buckets before resource-growth discovery becomes callable
• fail closed at execution time for empty or single-bucket history instead of returning a misleading successful trend
• expose a stable RESOURCE_HISTORY_UNAVAILABLE reason while retaining source-specific probe evidence privately

Root cause

Capability detection treated a successful metadata query as proof that historical evidence existed. On a real Doris 4.0.5 runtime, the audit source was permission denied and the partition source was readable but contained zero recorded time buckets, yet the historical child could still appear usable.

Verification

• uv run pytest -q -W error: 1830 passed, 85 skipped, 68.00% coverage
• focused capability, manifest, dispatcher, and runtime suite: 322 passed
• uv lock --check
• generated catalog check
• Ruff, Mypy, Bandit
• source distribution and wheel build

Real Doris boundary

Controlled loopback validation against Doris 4.0.5 used the existing read-only runtime identity. Modern HTTP 2026-07-28, Legacy HTTP 2025-06-18, and SDK STDIO 2025-11-25 all exposed 11 cluster children with analyze_resource_growth callable=false and reason RESOURCE_HISTORY_UNAVAILABLE. Exact calls returned CHILD_CAPABILITY_UNAVAILABLE and no historical payload. Protected runtimes remained ready and the isolated fixture was removed.

Security

The validation retained no credentials, raw SQL, native identifiers, or result rows. It performed read-only probes only and did not add an admin fallback.