C6 numerics: authenticate complete baseline and mutant reports (#334)

.github/workflows/c6-cluster-coefficients-numerics.yml

@@ -5,6 +5,8 @@ name: C6 cluster coefficient numerics (replay)
     paths:
       - 'frontiers/c6_cluster_coefficients_numerics_20260930/**'
       - '.github/workflows/c6-cluster-coefficients-numerics.yml'
+      - 'tools/c6_coefficient_numerics_replay.py'
+      - 'tests/test_c6_coefficient_numerics_workflow.py'
   workflow_dispatch:
 
 permissions:
@@ -21,6 +23,13 @@ jobs:
       - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97
         with:
           python-version: '3.11.16'
+      - name: Verify full-shell output contracts in both Python modes
+        env:
+          PYTHONDONTWRITEBYTECODE: '1'
+        run: |
+          set -euo pipefail
+          python -B -S -m unittest discover -s tests -p test_c6_coefficient_numerics_workflow.py -v
+          python -B -O -S -m unittest discover -s tests -p test_c6_coefficient_numerics_workflow.py -v
       - name: Verify manifest and main-resident pins, replay exact controls and GH40, reject mutants
         env:
           PYTHONDONTWRITEBYTECODE: '1'
@@ -48,15 +57,9 @@ jobs:
               blob = hashlib.sha1(b'blob %d\0' % len(data) + data).hexdigest()
               if len(data) != entry['bytes'] or hashlib.sha256(data).hexdigest() != entry['sha256'] or blob != entry['git_blob']:
                   raise ValueError('consumed source identity mismatch on this tree: ' + entry['path'])
-          script = str(root / 'coefficients.py')
-          for flags in (['-B', '-S'], ['-B', '-O', '-S']):
-              out = subprocess.run([sys.executable, *flags, script, '--check'], cwd=root, capture_output=True, timeout=900)
-              if out.returncode != 0:
-                  raise RuntimeError('check failed under ' + ' '.join(flags) + ': ' + out.stdout.decode()[-400:])
-              for m in ['cubic-sign', 'antiderivative', 'typed-boundary']:
-                  rc = subprocess.run([sys.executable, *flags, script, '--check', '--mutant', m], cwd=root, capture_output=True, timeout=900).returncode
-                  if rc != 1:
-                      raise RuntimeError('mutant ' + m + ' not rejected')
+          sys.path.insert(0, str(pathlib.Path('tools').resolve()))
+          from c6_coefficient_numerics_replay import replay
+          replay(root)
           print(json.dumps({'scientific_effect': 'NONE', 'certified': False}))
           PY
           git diff --exit-code

reviews/c6_numerics_protocol_20261006/SCOPE.md

@@ -0,0 +1,95 @@
+# C6 numerical-workflow report contracts — AUD-C6NUM-PROTOCOL-01
+
+Dylan Roy — delegated AI engineering. Author: OpenAI / GPT-6 Astra Pro,
+`qs-parent-c6num-repair-20261006`, pickup Math-#334/6009005443.
+Scientific effect NONE. Same-account/provider organizational-independence credit 0.
+This is an engineering repair, not a new mathematical or numerical acceptance.
+
+## Source and ownership
+
+Original source read at `516235290e12decdcd223cfb20928a2d97b824f1`:
+
+| Source | Git blob | Bytes | SHA256 |
+|---|---|---:|---|
+| workflow | `268ca3fb856b818022784f87bea96d0714c146aa` | 3180 | `06ee6f3beed34eb85ee45e0bbc19ce5e94aeab474a0dd31a492f48b93ca5f78c` |
+| coefficients.py | `bd45a9fdecf300111b646644384dcead20c2e529` | 18825 | `bcd63f0806a58a76a85189372bc6eee99fc3847c74dcc34906ef63464a64b97b` |
+| RESULTS.json | `2648314f8c11f0637ce301ba76360142a8309e8f` | 4095 | `9c88ea85b6ba08199996c9b5ff74ecd06afc0f458e08c28b4c3af6191f437fcb` |
+
+The original finding belongs to `pr321-followthrough-wrapper-audit-20261006`.
+Its report and later authentic-characterization contribution #334/6009064997
+remain separately attributed. A crossed later writer pickup6009007450 was
+acknowledged in6009140749; earlier6009005443 retains the single production route.
+No other writer's implementation or source review is imported here.
+
+## Measured fixed contracts
+
+Local CPython 3.13.5, unchanged checker and RESULTS, both `-B -S` and `-B -O -S`:
+
+| Invocation suffix | Exit | Exact structural report | Stdout bytes / SHA256 |
+|---|---:|---|---|
+| --check | 0 | `{"mode":"check","passed":true,"scientific_effect":"NONE"}` | 63 / `c54d375daf9298c6a7cc591191dd9e3508b80213ddd24e5320859b429acefff6` |
+| --check --mutant cubic-sign | 1 | `{"error":"classifier agrees with the direct critical-point count on the typed domain","passed":false}` | 105 / `b0271f45b76f6b28702624c2e6efce69bbd54e1d2246fdca3d09aa553d9095b0` |
+| --check --mutant antiderivative | 1 | `{"error":"[LM] cubic mass 48 k^5 on n = 2","passed":false}` | 62 / `2f60570838650467fb3deeadf8138cf1809f355da1db0511e68cabca7f263ad5` |
+| --check --mutant typed-boundary | 1 | Same classifier report as cubic-sign | 105 / `b0271f45b76f6b28702624c2e6efce69bbd54e1d2246fdca3d09aa553d9095b0` |
+
+All eight original characterization processes completed with empty stderr;
+all four mode pairs were byte-identical. Actual original stdout uses sorted
+keys, default JSON spaces and one trailing newline; the table compacts spacing.
+The validator intentionally allows harmless JSON whitespace/key ordering but
+requires the complete key set, exact Boolean/string types and exact values.
+It rejects duplicate keys, numeric values, nonfinite values, extra or missing
+fields, trailing documents, malformed UTF-8, silence, wrong reports, wrong
+exits, signals and unexpected stderr. Expectations are fixed source, never
+learned from the invocation under test. Timeout/launch exceptions propagate.
+
+The two classifier mutants really share one diagnostic. Output validation
+cannot distinguish a malicious replacement emitting that same report. The
+unchanged source/membership and consumed-pin guards remain essential. The
+helper is not a sandbox or a replacement for those guards.
+
+## Test-first execution and preserved workflow
+
+Twelve full-shell regression methods execute 132 distinct committed synthetic
+Git-repository scenarios. Each fixture runs the actual named replay Bash block,
+including manifest/source/current-pin checks and final `git diff --exit-code`.
+The trace is outside the checkout; children are real processes, not mocks.
+
+Against the ORIGINAL workflow: **100 intended assertion failures, zero errors**;
+109 fixtures exit 0 (100 invalid, nine valid/formatting), 23 reject as intended;
+952 child invocations. Against the repair: all 132 scenarios behave correctly,
+both normal and optimized outer test runs; each run has 594 child invocations,
+nine valid/formatting admissions and 123 rejections. These are 132 distinct
+scenarios, not 264 kinds of evidence. They are not scientific checker results.
+
+The genuine repaired helper then executes all eight original checker invocations
+with the same 900-second child budget. It exits 0 with empty stderr; every
+reported output hash/exit equals the independent characterization. The full
+numerical packet's source-manifest shell was NOT replayed locally: only the
+three authenticated original files and engineering sources were available.
+Hosted dedicated execution must verify the actual complete repository tree.
+No no-argument generation, Monte Carlo rerun, RESULTS edit, or tolerance change
+occurred. The original numerical comparisons remain noncertified.
+
+The workflow's complete manifest/pin preflight and final JSON/Git postflight
+are byte-identical. Only the intervening exit-only loop is replaced by the
+helper; two focused test invocations and their trigger paths are added. Runtime
+3.11.16, action pins, permissions, child 900 seconds and job 20 minutes are
+unchanged. Fixtures isolate their own Git settings and disable automatic
+maintenance; this is not a repair claim about #321's separate cleanup race.
+
+Python's primary reference contracts are `subprocess.CompletedProcess` and
+`json.loads`/`object_pairs_hook`/`parse_constant` in the official Python 3.11
+documentation. Numeric and duplicate rejection is explicit, not reliance on
+default JSON parsing. No third-party package is introduced.
+
+## Remaining review and integration requirements
+
+This four-file proposal changes no byte in the numerical packet, mathematical
+sources, RESULTS, manifests, source maps, other workflows, formal gate, or
+scientific-status surfaces. A separate nonauthor engineering review, actual
+pinned-runtime dedicated/general checks and eligible expected-head integration
+are required. Author local execution is not a review or merge receipt. Old
+failed records remain valid historical failures. No full repository local
+suite, local Lean execution, independent whole-program alignment or scientific
+closure is claimed. Retained in-memory/printed diagnostics do not guarantee raw
+child-output persistence after process or filesystem failure.