[surlymochan]

Adds a reproducible head-to-head benchmark against obra/superpowers v7.0.0 and hardens the TDD gates the benchmark found gaps in (released as 0.2.1).

Gate fixes (general checks, each with unit tests):

• Verifiable proofs: I6a binds command, exit code, output hash, file hashes and HEAD, and snapshots the tree. I6b re-validates the binding and re-runs the command in the snapshot.
• Red and green must target the same tests.
• skip/todo tests don't count as red or green.
• Changed tests must reference changed symbols, and type-only or existence-only assertions are flagged when they are the only ones.
• Python assertions are now recognised.
• J4b fails contradictory MUST/SHALL requirements.
• New tdd_mode values refactor and docs, a legitimate exemption with misuse checks.

Benchmark (bench/):

• 24 original and 12 held-out scenarios. The held-out ones were written by a separate author after the fixes, without access to the gate code.
• Results: v0.2.0 caught 10/17 faulty scenarios with 3/7 false positives. v0.2.1 caught 16/17 original, 6/8 held-out, 0/11 false positives.
• LLM judge arms with grok-4.7-high, one run per scenario: superpowers skills 25/25 with 1/11 false positives; no skills 25/25 with 0/11.

End-to-end (bench/e2e/): 8 tasks with hidden tests, same model for every arm.

Arm	Success	Median time	Median tokens	n
baseline	24/24	55 s	105k	24
superpowers	24/24	187 s	312k	24
OpenFlow	11/12	985 s	5.29M	12

The README now opens with these tables, losses included.