feat(power): typed adversarial engineering verification

[Kitahl]

Scope

Power-only vNext upgrade, security hardening, and frozen old-vs-new Power benchmark stacked on exact validated Space head 01c07faf1848284bda3c13d1c1eec972629be9c4.

Reality is not included or started.

Power implementation

Power vNext adds content-bound failure hypotheses; typed direct/regression/real-entrypoint/negative-control/mutation/property/metamorphic/differential/environment checks; substantial-change minimum verification; residual failure coverage; mutation-survival blocking; artifact/harness/environment/oracle diagnosis; neutral dual-verifier repair admission; and historical receipt compatibility.

Final security review also hardened the new python-script verifier: a script must resolve inside the supplied repository root. Absolute or symlink-resolved escapes are UNAVAILABLE and are not executed. shell=False, active-interpreter binding, trusted executable resolution, per-check timeout handling, disabled custom commands by default, and hashed stdout/stderr remain preserved.

Final published revision

• branch: upgrade/challenge-vnext-power
• stacked base: upgrade/challenge-vnext-space
• base SHA: 01c07faf1848284bda3c13d1c1eec972629be9c4
• final head SHA: 09b6746e68df0669ba08161917467d3f916f2a93
• final tree SHA: 00ea98d5e23c58f91997a528d8082d03508ab47a
• commits over Space: 10
• changed paths over Space: 9
• final Power runtime blob: 99f5b955b782b61ccaa5fa481ecd347963c3a35a

The benchmark commits do not modify tools/power_runtime.py; the runtime blob remains the same hardened Power implementation previously validated before benchmark instrumentation.

Final changed paths

1. tools/power_runtime.py
2. tests/test_power_challenge_runtime.py
3. tests/test_power_entrypoint_security.py
4. docs/specs/POWER_ENGINEERING_SPEC.md
5. skills/codebot/SKILL.md
6. benchmarks/POWER_VNEXT_AB_PROTOCOL.md
7. benchmarks/fixtures/power_runtime_space_base.py
8. benchmarks/harness/bench_power_vnext_ab.py
9. tests/test_power_vnext_ab_benchmark.py

Frozen old-vs-new benchmark

Benchmark type: deterministic mechanism-conformance and adversarial-discrimination. It is not a benchmark of general software correctness, repair efficacy, or production reliability.

Pinned implementations:

• OLD Space-base revision: 01c07faf1848284bda3c13d1c1eec972629be9c4
• OLD Power runtime Git blob: 5b2c0e6f06df99bac77973f70485cd3c465729e4
• NEW Power implementation revision: b8e6557253a642ccc85d27a22c79241256eb3f9b
• NEW Power runtime Git blob: 99f5b955b782b61ccaa5fa481ecd347963c3a35a

The benchmark fixture points directly to the historical OLD Git blob. The harness recomputes both Git blob identities and aborts on drift.

Score

Metric	OLD	NEW	Delta
Shared controls	9/10 (90.0%)	10/10 (100.0%)	+1
vNext discriminators	0/12 (0.0%)	12/12 (100.0%)	+12
Total	9/22 (40.9%)	22/22 (100.0%)	+13 passes / +59.1 pp

Exact CI summary:

POWER_AB_SUMMARY={"delta_passes": 13, "new_shared": {"passed": 10, "total": 10}, "new_total": {"passed": 22, "total": 22}, "new_vnext": {"passed": 12, "total": 12}, "old_shared": {"passed": 9, "total": 10}, "old_total": {"passed": 9, "total": 22}, "old_vnext": {"passed": 0, "total": 12}}

Shared behavior defect found by A/B

The OLD implementation failed shared.unavailable_dominates_unknown: with one mandatory timeout and one mandatory unavailable verifier, OLD returned UNKNOWN. NEW returns UNAVAILABLE, so missing mandatory verifier capability is no longer masked by an inconclusive timeout.

vNext discriminators NEW passed 12/12

1. duplicate semantic failure hypotheses rejected;
2. task/binding mismatch fails closed;
3. substantial-change minimum rejects omitted regression verification;
4. killed mutation records discriminator success;
5. surviving mutation blocks “fixed”;
6. real in-repository Python entrypoint executes;
7. outside-root Python entrypoint is refused and not executed;
8. metamorphic failure is relation-scoped;
9. TASK_ARTIFACT and AGENT_HARNESS remain distinct diagnostic candidates;
10. oracle changes change evidence identity;
11. repair self-certification is rejected;
12. independent dual verification may become COMMITTABLE while execution_authorized=false and host commit remains required.

OLD receives no synthetic passes for mechanisms it does not implement.

Final benchmark-head validation

Exact final benchmark head 09b6746e68df0669ba08161917467d3f916f2a93 / tree 00ea98d5e23c58f91997a528d8082d03508ab47a:

• Research software validation run 33238871207: PASS
  • Ruff: PASS (All checks passed!)
  • full repository suite: 632/632 PASS (Ran 632 tests in 33.488s, OK)
  • old-vs-new benchmark test: PASS
  • Soul/Gauntlet public invariant validator: PASS
  • typed-runtime validator: 52/52 PASS
  • showcase validator: 33/33 PASS
  • compileall/syntax: PASS
• Runtime portability run 33238871178: PASS
  • Ubuntu: PASS
  • macOS: PASS
  • Windows: PASS
  • stable portability gate: PASS
• Security gates run 33238871177: PASS
• CodeQL run 33238871183: PASS
• Existing FOIL benchmark pilot run 33238871192: PASS

Authority and claim boundary

Power emits claim-scoped ENGINEERING evidence only. execution_authorized=false; no repair self-promotion, host write, or merge authority is added. Soul retains automatic routing/release authority and the neutral repair gate remains host-controlled.

The A/B result establishes only performance on the 22 frozen mechanical cases. It does not establish test-suite completeness, semantic correctness of every oracle, mutation score as a production-reliability proxy, cross-repository repair efficacy, exhaustive correctness from passing metamorphic relations, or general real-world software correctness.

Completion state

Power is implemented, security-hardened, benchmarked old-vs-new, fully validated at the final published head/tree, and remains open as PR #58 stacked on Space. main was not merged. Reality was not started.