Reference genuineness: the reference report is the file product of running
this exact workflow. To reproduce from scratch (Python 3.11+ stdlib, ~seconds):
  1. python3 contributor_kit/generator.py          (rebuilds inputs + ground truth from seed)
  2. python3 contributor_kit/reference_solver.py   (recomputes reference/reconciliation_report.json)
  3. python3 contributor_kit/inferability_check.py (proves recoverability from agent-visible data)
  4. python3 evaluation/checker.py <candidate> reference/reconciliation_report.json  (grader; exit 0=pass)
Step 1 regenerates, byte-identically, every data file in the separately
uploaded input bundle — diff them to confirm the reference was computed from
exactly those inputs. README.md is the authored task brief and ships in the
input bundle.
Verification order: run `python3 freeze_manifest.py verify .` FIRST, on the
pristine extraction (it is strict: files created by reproduction count as
additions). After steps 1-3, confirm genuineness by checking that the
regenerated reference still matches its frozen hash in FREEZE_MANIFEST.json.
