Add systematic result metadata summarization

[tadamcz]

Summary

• add Inspect tasks for Lite sequence, conjecture, and accepted-proof summaries
• preserve the interactive, network-isolated Lean sandbox for proof summarization
• collect normalized root catalogs and per-sample metadata deterministically from eval logs
• export the root metadata directory alongside selected plaintext runs
• add fast regression coverage for current dataset loading and the pinned sandbox image

Validation

• ruff check on all changed Python files
• strict mypy on all changed Python files
• 28 focused dataset, registry, pin, and summarization tests
• constructed 99 sequence, 100 conjecture, and 44 Fable proof samples
• regenerated Lite catalogs and 44 Fable metadata files in a temporary directory; outputs matched the existing generated files byte for byte
• exporter dry-run against the results repository reported no changes

Scope

No generated metadata, logs, plaintext runs, or exported results are included in this PR.

Compatibility fixes found while rebasing onto main

The original results-branch implementation referenced the pre-reorganization OEIS metadata paths and older load_subset and image-tag APIs. This branch uses the canonical metadata/snapshots and metadata/derived paths, passes OEIS_DIR to load_subset, and pins the sandbox image build to the OEIS Formal Conjectures commit.

A model-backed Inspect eval and Docker rebuild were not rerun for this PR; task construction and compose generation were verified without starting Docker.