[P2] Add memento-vault benchmark runner for Rondo harness evaluation #29

Context

The handoff's proving ground is memento-vault. The evaluation plan calls for running 8-12 real issues and measuring whether externalized knowledge/process reduces human steering and review failures.

Blocked by

• memento-vault benchmark task definitions.
• Rondo run ledger/checkpoints.
• Rondo adapter contract.
• Rondo gate runner.
• GitHub Issues tracker support for the pilot repo.

Proposal

Add a benchmark runner that consumes curated memento-vault tasks and records comparable run outcomes.

Scope

• Load benchmark task definitions with issue snapshot, acceptance criteria, required gates, risky areas, and rubric.
• Run one or more Rondo adapters against the task set.
• Collect metrics: interventions, gate failures, repeated mistakes, PR-readiness, time, and harness failures.
• Classify failures using the taxonomy from the evaluation plan.
• Produce a batch report after 3-5 tasks.

Acceptance criteria

• A benchmark batch can run against a configured task set without manual result collation.
• Results distinguish agent failure, harness failure, gate missing, knowledge missing/not retrieved, and ambiguous requirements.
• Batch reports identify repeated failures that should become guides, gates, or policy changes.
• The runner is adapter-neutral where practical.