[codex] add benchmark runner and baseline

[divo12]

Stack

Depends on #7. Review against codex/mine-exports-ledger.

Problem

Exported eval suites are immutable manifests, not proof that the champion is reproducible. OFW needs one runner that executes only developer-visible cases under the frozen runtime/verifiers, establishes a durable baseline, detects semantic drift, labels synthetic cases, and stops before exceeding hard work budgets.

Implementation

• Adds BenchmarkRunner(processed_harness, export_bundle, policy).
• Verifies harness revision and execution/lifecycle fingerprints match the benchmark manifest.
• Rejects selection/admission cases if they are injected into the developer suite before environment preparation.
• Requires trace family, cluster family, and partition to match the authoritative ledger, so relabelled holdouts also fail.
• Requires the exact snapshot path and digest to match the ledger, so authorized labels cannot swap in holdout payloads.
• Verifies every snapshot path stays under the harness artifact root and matches its content digest.
• Runs lifecycle and all independent verifiers with reset between attempts and guaranteed cleanup.
• Records verifier errors/abstentions as explicit failed attempts; no score is silently dropped.
• Supports deterministic repeats and metadata-safe whitespace/noise simulation.
• Labels every synthetic attempt and caps its benchmark weight.
• Enforces a hard maximum attempt budget and returns an explicit partial status.
• Computes semantic digests without runtime-duration noise.
• Writes durable baseline and benchmark result artifacts atomically.
• Re-runs the champion and aborts on output/verdict/policy/runtime drift.

TDD evidence

• Stable champion baseline reproduces across repeated runs.
• Selection/admission cases never execute through the developer runner.
• Relabelling a holdout as frontier still fails the ledger check.
• Swapping a holdout snapshot under authorized developer labels still fails.
• Reset failure returns ENVIRONMENT_ERROR and persists completed attempt evidence.
• Unstable output triggers BASELINE_DRIFT.
• Python verifier errors remain visible failures with zero pass rate.
• Repeats and synthetic attempts are labeled and weighted correctly.
• Hard attempt budget stops at the exact limit with BUDGET_EXHAUSTED.
• Baseline artifact is persisted.

Verification

• pytest -q — 109 passed
• Ruff and strict mypy across src and tests — clean
• forbidden construct scan — no Any, raw dict, object, getattr, setattr, lambdas, or __import__
• Bandit and deep security scan — zero findings
• pip-audit — no known vulnerabilities
• source distribution and wheel build successfully

Deliberate scope

• Hard v0 budget is attempt count; cost/token budgets require lifecycle cost telemetry rather than invented estimates.
• Simulation adds deterministic formatting/noise only to privacy-safe snapshots; richer user-text mutation waits for an explicit I/O read policy.
• Candidate comparison, selection/admission execution, mutation, scheduler, and promotion are not included.