[codex] add no-leak Mine exports

[divo12]

Stack

Depends on #6. Review against codex/mine-diagnosis-clustering.

Problem

Admitted traces and failure clusters cannot be exported independently: exact/template families would leak between training, memory, regression, selection, and admission. OFW needs one authoritative family ledger before materializing any downstream artifact.

Implementation

• Adds MineExports(revision, mine_result, diagnosis_result, policy).run().
• Derives conservative trace families from normalized privacy-safe observation topology.
• Canonicalizes topology order so reordered copies remain one family.
• Uses stable cluster family IDs from diagnosis revisions.
• Assigns every eligible trace/cluster family exactly one typed partition.
• Carries previous ledger assignments forward and fails closed if a family flips good/failure role.
• Supports explicit typed cluster partition rules; unlisted clusters use stable family hashing, never sorted position.
• Produces a GoodTraceDataset of safe content-addressed snapshot references with family-consistent train/validation splits.
• Requires approved consent, explicit license, and metadata-only privacy transform provenance.
• Produces separate developer frontier/regression, selection holdout, and admission holdout suites.
• Produces immutable memory proposals without changing harness files.
• Produces a benchmark manifest bound to harness execution/lifecycle fingerprints.
• Writes owner-only atomic ledger, dataset, suite, memory, benchmark, and bundle artifacts.

TDD evidence

• Same-topology traces share one family assignment and one dataset split.
• Training families never appear in developer or holdout suites.
• Selection/admission artifacts are physically separate from developer-visible suites.
• Same family requested for training and eval fails closed.
• Memory partition creates a proposal while leaving the harness unchanged.
• Adding a lexically earlier cluster cannot reassign existing cluster families.
• Identical inputs produce identical content-addressed bundles.
• Real Mine-produced snapshots pass digest verification; writer and reader hash identical bytes.
• Same topology in different observation order cannot cross training and eval.

Verification

• pytest -q — 101 passed; latest full coverage remains 93%
• Ruff and strict mypy across src and tests — clean
• forbidden construct scan — no Any, raw dict, object, getattr, setattr, lambdas, or __import__
• Bandit — zero findings
• pip-audit — no known vulnerabilities
• source distribution and wheel build successfully

Deliberate scope

• Exports reference privacy-safe snapshots; raw Langfuse I/O hydration remains an explicit later read-policy surface.
• Benchmark execution, candidate mutation, A/B gates, scheduler, and promotion are not included.
• Conservative topology families may over-group; loosen only with reviewed family evidence, never by silently splitting holdouts.