Skip to content

Enforce state budgets, post-death run_late parity, must_fail gate, and LINT_PASS (#53 A3/A4/A6/A7) - #60

Merged
zero-aska merged 5 commits into
mainfrom
feat/verification-discipline-a3-a4-a6-a7
Nov 19, 2026
Merged

zero-aska merged 5 commits into
mainfrom
feat/verification-discipline-a3-a4-a6-a7

Conversation

@ahrazzle

Copy link
Copy Markdown
Owner

Close the remaining open verification-discipline items from #53 (A3 test_07 wall-clock budget, A4 post-death native run_late parity, A6 must_fail certificate controls and tier-swap gate, and A7 static checker-auditor LINT_PASS status) without touching search results or formal_kernel.py.

What Changed

  • A3 (Deterministic state budget in soko_optimize_route and test_07): Added optional per-stage max_states cap (budget_sec=0 disables wall-clock cutoffs) in cpp/soko_engine.cpp, soko_mcp/engine_bridge.py, and soko_mcp/server.py. test_07 now reproduces III-20 87 -> 84 via max_states=6000000, budget_sec=0, and soko_engine.cpp uses standard C++20 headers for GCC and Clang compatibility.
  • A4 (Post-death run_late parity in cpp/soko_engine.cpp): Replaced early hero-death returns in run_late with OBJ_CORPSE ('X') so simultaneous monster kills, door transitions, monster gates, and Bard notes match PuzzleScriptOracle._run_late. Enabled unconditional final_ascii checks across all 528 sampled catalog routes in tests/test_engine_differential.py (test_perft.py and test_bench.py stay intact).
  • A6 (must_fail controls and tier-swap gate): Added must_fail and verify_certificate_gate in soko_mcp/evidence.py, recorded live dual-oracle and gate must_fail_controls plus certificate_sha256 digests in engine_bridge.py and tracks/*/verify.py, and added tests/test_certificate_corruptions.py catching missing controls, hash tampering, tautological seeded controls, and LOWER_BOUND / LINT_PASS -> VERIFIED tier swaps.
  • A7 (LINT_PASS on static checker audits): Relabeled audit_dual_checkers and audit_builtin_soko_oracles in soko_mcp/checker_auditor.py from VERIFIED to LINT_PASS, and refreshed SHA256SUMS via python3 -m soko_mcp.integrity --write.

What Is Not Proved

LINT_PASS checks static AST separation and exact-math syntax rather than semantic equivalence between checkers. LOWER_BOUND_UNWITNESSED refutes shorter routes under the engine's runtime-checked heuristic without proving a winning route exists at that bound.

Verification

  • python3 -m soko_mcp.integrity (21/21 locked sources verified)
  • mde guard --code-audit (0 findings across all 10 touched Python modules and tests)
  • python3 -O -B -m pytest across test_certificate_corruptions.py, test_v04_foundation0_synthesis.py, test_mcp_server.py, test_engine_differential.py, test_perft.py, test_bench.py, test_f04_close.py, test_optimize_seed.py, and test_integrity.py (54 passed, 612 subtests passed)

Automated posting by agentic team with human oversight.

Integrate merged main (#57, #58, #59, #20) while preserving verification
discipline gates, must_fail controls, and post-death corpse-render parity.
Regenerate SHA256SUMS via python3 -m soko_mcp.integrity --write.

Co-authored-by: ahraz <ahrazzle@users.noreply.github.com>
cursor Bot pushed a commit that referenced this pull request Nov 19, 2026
Resolve SHA256SUMS against #60 corpse parity and gate changes; keep
PR #60 cpp/soko_engine.cpp semantics. Mark soko_eval_live_receipt.json
as historical (382e9da) and document II-4 mismatch as pre-TCB-fix.

Co-authored-by: ahraz <ahrazzle@users.noreply.github.com>

@zero-aska zero-aska left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Approve at 537e59e. Cloud review bc-2e58906b, run 2.

This approves the gate behaviour and the corpse-render parity fix from the earlier review, now with main (3ea918f) merged in normally, with no force-push. The only conflict was SHA256SUMS, which was regenerated with soko_mcp.integrity --write.

Check Head main 3ea918f
soko_mcp.integrity PASS 21/21
Targeted pytest 54/54
unittest discover 307 ran, 3 failures, 20 errors 297 ran, 3 failures, 22 errors

The 3 failures are the same C++ landmark/transport compile failures as on main, and the errors come from optional experiments/ imports. This head adds no new failures.

@zero-aska
zero-aska merged commit fc7018d into main Nov 19, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants