Add autonomous contract and scale evaluation

[kamjcx]

Scope

Developer-only evaluation delivery. This PR does not modify src/a1_factor_engine/, retrieval, ranking, qualification, factor IDs/values, formal catalogs, or existing frozen answers.

Added

• independent contract Oracle and 414 deterministic public-synthetic cases
• exact boundary/subject/unit/provenance/source-priority checks
• metamorphic, high-risk-neighbour and eight workflow state-machine attacks
• explicit-denominator metrics, Bad Case attribution, immutable first-run JSON/Markdown/SHA evidence
• 10k/50k scale benchmark with c10/c25/c50 latency, RSS, throughput and stability
• CI evidence execution and public methodology/limitations

Immutable first run

The frozen evaluator commit fe4eee7 produced 314/418 complete-row passes and an explicit NO-GO. Top-1 93.05%, Recall@5 99.23%, abstention 90%, MORE_INPUT 20%, evidence metadata 93.27%, replay 100%, zero boundary/subject/unit/HTTP-500 violations, six geography/year forbidden escapes, and 7/8 workflow attacks. All 104 Bad Cases remain unchanged.

The original Manifest git_dirty timing defect is preserved and explained by an erratum; it did not alter raw results.

Performance

10k/50k stability gates pass. At 50k: 117.67s index build, 1.409s warm P95, 205.5MB peak RSS, ~0.75 req/s one-process throughput. This is descriptive synthetic evidence, not an SLA.

Local QA

• Python 3.11/3.12/3.13: 350 passed each
• core branch coverage: 87.06%
• autonomous focused: 26 passed; strict mypy/Ruff/compileall pass
• existing FactorBench V1/V2/V3, Frozen Unit, Sealed Unit v2/v4, Portfolio Challenge: pass
• package build/archive isolation: pass
• local Docker daemon unavailable; remote container check is authoritative

Runtime defects, if adjudicated, will be repaired only in a separate PR after this evidence PR.