d6g8k5htny-coder · Oct 6, 2026 · Oct 6, 2026 · Oct 6, 2026
@@ -5,6 +5,7 @@ name: Soft fold limit exact controls
     paths:
       - 'frontiers/soft_fold_limit_20261002/**'
       - '.github/workflows/soft-fold-limit.yml'
+      - 'tests/test_soft_fold_workflow.py'
       # bound merged sources (consumed and cited in SOURCES.json); the cited unmerged sources are bound through the API
       - 'frontiers/three_fronts_20260924/LIFETIME_REMAINDER.md'
       - 'imports/lifetime_parent_20260925/UNIFORM_MATRIX_CAP_AND_LIFETIME.md'
@@ -37,6 +38,13 @@ jobs:
       - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97
         with:
           python-version: '3.11.16'
+      - name: Exercise complete-shell output contracts in both modes
+        env:
+          PYTHONDONTWRITEBYTECODE: '1'
+        run: |
+          set -euo pipefail
+          python -B -S -m unittest discover -s tests -p test_soft_fold_workflow.py -v
+          python -B -O -S -m unittest discover -s tests -p test_soft_fold_workflow.py -v
       - name: Verify sources, replay both modes, reject mutants
         env:
           PYTHONDONTWRITEBYTECODE: '1'
@@ -150,17 +158,33 @@ jobs:
                                     'via': where, 'current_ref_checked': now, 'dependency_merged': merged}))
           script = str(root / 'fold_check.py')
           expected = (root / 'RESULTS.json').read_bytes()
+          # Fixed native carriers measured at original checker c2e1e77e in run37464626067.
+          failures = {
+              'M1': b'mutant M1: failing controls F1_model_pins_and_weight\n',
+              'M2': b'mutant M2: failing controls F2_lemma_FL2_generic\n',
+              'M3': b'mutant M3: failing controls F3_lemma_FL1_window\n',
+              'M4': b'mutant M4: failing controls F4_proposition_FL4_certificates\n',
+              'M5': b'mutant M5: failing controls F5_theorem_FL_bookkeeping\n',
+              'M6': b'mutant M6: failing controls F6_corollaries\n',
+              'M7': b'mutant M7: failing controls F4_proposition_FL4_certificates\n',
+              'M8': b'mutant M8: failing controls F7_proposition_FL7_identification\n',
+          }
+          inventory = {'M1': 'F1', 'M2': 'F2', 'M3': 'F3', 'M4': 'F4',
+                       'M5': 'F5', 'M6': 'F6', 'M7': 'F4', 'M8': 'F7'}
+          if (type(manifest['mutants']) is not dict or list(manifest['mutants']) != list(inventory)
+                  or manifest['mutants'] != inventory):
+              raise RuntimeError('complete ordered M1-M8 control mapping required')
+          stages = [('baseline', [], 0, expected, b'')]
+          stages += [(m, ['--mutant', m], 1, b'', carrier) for m, carrier in failures.items()]
+          stages += [('M99', ['--mutant', 'M99'], 2, b'', b'unknown mutant label\n')]
           for flags in (['-B', '-S'], ['-B', '-O', '-S']):
-              out = subprocess.run([sys.executable, *flags, script], check=True, capture_output=True, timeout=600).stdout
-              if out != expected:
-                  raise RuntimeError('output differs from RESULTS.json under ' + ' '.join(flags))
-              for m in manifest['mutants']:
-                  rc = subprocess.run([sys.executable, *flags, script, '--mutant', m], capture_output=True, timeout=600).returncode
-                  if rc != 1:
-                      raise RuntimeError('mutant ' + m + ' not rejected')
-              rc = subprocess.run([sys.executable, *flags, script, '--mutant', 'M99'], capture_output=True, timeout=600).returncode
-              if rc != 2:
-                  raise RuntimeError('unknown mutant label not rejected with exit 2')
+              for label, args, code, stdout, stderr in stages:
+                  result = subprocess.run([sys.executable, *flags, script, *args], capture_output=True, timeout=600)
+                  if (result.returncode, result.stdout, result.stderr) != (code, stdout, stderr):
+                      raise RuntimeError(label + ': complete exit/stdout/stderr contract differs under ' + ' '.join(flags))
+                  print(json.dumps({'stage': label, 'flags': flags, 'returncode': result.returncode,
+                                    'stdout_sha256': hashlib.sha256(result.stdout).hexdigest(),
+                                    'stderr_sha256': hashlib.sha256(result.stderr).hexdigest()}, sort_keys=True), flush=True)
           print(json.dumps({'scientific_effect': 'NONE', 'mathematical_acceptance': False}))
           PY
           git diff --exit-code
@@ -0,0 +1,151 @@
+# Soft-fold workflow output contract
+
+Dylan Roy — delegated AI engineering. Actual author OpenAI / GPT-6 Astra Pro,
+`soft-fold-protocol-followthrough-20261006`; #322 pickup6016299453.
+Scientific effect NONE; organizational-independence credit0. No author self-merge.
+
+## Original sources and independently measured carriers
+
+Source cut11f9127f00ce922e5d39b0da42af76134eb78a21. Original workflow
+`.github/workflows/soft-fold-limit.yml`:10578 bytes, Git blob
+839b99a47a4b9d0a79525216118717b22bad4614, SHA256
+0aec5cb657a2fa37bab5f77f4f77ca6f978fbd9297a0b56b691b0f93d6ec2c6f.
+Checker `frontiers/soft_fold_limit_20261002/fold_check.py` is unchanged blob
+c2e1e77e9f16bc7c62fbb855b55dc2f72e5a8cc2. RESULTS is unchanged blob
+604c53fadd71312baa3b4bfdbbf1760eddc4ac66.
+
+An observation-only draft predecessor e710ba569feb5d66f5bac282de689758a9b22111
+first obtained actual pinned-runtime evidence. Original run37464626067/1,
+job112272450916, checked out64831bd92336733780928dc40e55db5304025123
+(parents11f9127f+e710ba56), installed CPython3.11.16/Git2.55.0, authenticated
+both workflow and checker, executed the exact original complete source preflight,
+and ran20 authentic processes. Complete original log read; selected observations
+are recorded in PR370/6016447033. No local authentic checker replay is claimed:
+Git clone failed DNS and direct source download was unavailable.
+
+Both baseline invocations exit0 with2903-byte frozen RESULTS stdout, SHA256
+459bbdf3c84b77548e53824df889b13a8ebd0fe6b0d7d7cadf8a4d73821641e9,
+and empty stderr. All eight named mutants exit1 with empty stdout and exactly:
+
+`mutant LABEL: failing controls CONTROL\n`
+
+|Label|CONTROL|
+|---|---|
+|M1|F1_model_pins_and_weight|
+|M2|F2_lemma_FL2_generic|
+|M3|F3_lemma_FL1_window|
+|M4|F4_proposition_FL4_certificates|
+|M5|F5_theorem_FL_bookkeeping|
+|M6|F6_corollaries|
+|M7|F4_proposition_FL4_certificates|
+|M8|F7_proposition_FL7_identification|
+
+M4 and M7 name the same control, but their complete label-bearing lines differ.
+M99 exits2, empty stdout, exactly `unknown mutant label\n` on stderr.
+All ten corresponding stdout/stderr identities match between interpreter modes.
+The native main function reports the sorted complete failing-control set, so
+an extra failed control is not accepted as the intended single failure.
+Observation success alone did not establish a correct validator. Its temporary
+workflow is removed from the final proposed tree, retained only in history.
+
+## Repair and preservation
+
+Only the existing inline process loop changes behavior. Require the exact native
+ordered M1-M8 control mapping before children; construct the fixed20-command inventory; compare each
+actual exit/stdout/stderr tuple with its independently measured complete contract.
+The baseline retains byte-exact RESULTS and now requires empty stderr. Mutants
+and M99 require empty stdout and their entire stderr, including the newline.
+Whitespace changes are deliberately rejected for these deterministic text CLIs.
+No expectations come from the child under validation; no scientific checker or
+RESULTS generation, parser adapter, extra production helper/import is introduced.
+Launch/timeout exceptions propagate; cwd, capture_output=True,600-second child
+budget, both original modes and command arguments are unchanged. Accepted calls
+print stream identities. Failure raw-output durability is not newly guaranteed.
+
+The7599-byte source preflight remains byte-identical, SHA256
+2ad437f73e5f0c27482dd272144deadc2959d85e946c40f0226c51ce6abe5211.
+The137-byte final JSON/Git postflight remains byte-identical, SHA256
+752aa03685dbbe54ec40c3e91b8709eb6453c41c146e9636b295b2129c586cdc.
+Sizes include YAML indentation. All packet, proof, manifest, result and consumed
+or cited dependency bytes remain unchanged. So do action/runtime pins, read-only
+permissions and ten-minute job budget. One test trigger and two focused discovery
+commands are added. Existing D5/Palm/factorial tests/** triggers provide their
+separate complete repository discovery on the actual candidate.
+
+## Test-first local execution
+
+Actual CPython3.13.5/Git2.47.3; nine unittest methods,147 distinct committed real-Git
+whole-shell scenarios. Full uninterrupted original command:111 intended assertion
+failures,0 errors (132.243s). Breakdown:96 named-mutant false admissions,2 baseline
+stderr false admissions,8 unknown-label false admissions,4 malformed-inventory
+admissions,1 missing-wiring assertion. One valid admission and36 correct rejecting
+controls remain. The original executes2504 child commands.
+
+The unchanged final test file runs against the repaired workflow:9/9 methods,
+147/147 expected outcomes in EACH normal/optimized outer mode (65.548s/65.565s).
+Each has one valid admission and146 rejections,1444 actual children. No skipped
+or partitioned cases and no subprocess mocks. Across the three completed runs:
+441 fixture records,882 raw shell streams and5392 child invocations. These are
+repetitions of147 scenario types, not441 distinct defects.
+
+All eight mutant stages and both child modes are targeted. Every protocol
+rejection must reach and stop at its exact ordered target; nine source/tree
+controls and four malformed inventories must stop before any child. The final
+tracked-output mutation reaches all20 invocations before Git rejects it. Real
+Git configuration is isolated to temporary environments/commands; no user signing,
+hooks or maintenance settings are changed. Synthetic baseline/carriers never
+import expectations from production code.
+
+The local fixtures have nonempty tracked consumed and cited source pins. Their
+remote-source arrays are empty; no live API replay is claimed locally. The
+original hosted preflight actually resolved consumed #242 as merged, checked
+current ref54c6759e8434846941eb5218601dfef69c2c9d98 and source blob
+271412dbc96a29590f5f805258e15c9e335cb430 via TREE. The unchanged historical
+head/path/blob fallback was not exercised by that observation. The final dedicated
+workflow must supply its own actual full-packet candidate execution.
+
+## Required author correction after the first candidate failed
+
+Initial published head77b1b76f1c0b4a62e2359e83903c7b86a0c0db53 had a wrong
+list-only inventory assumption, mirrored in its synthetic fixtures. Actual
+SOURCES.json blob37b8a38a42a2dddbc4b2fdfdbae3a9cd4e28d7a6 stores an ordered
+object: M1:F1,M2:F2,M3:F3,M4:F4,M5:F5,M6:F6,M7:F4,M8:F7. The observation
+loop had correctly expanded its keys, but the first implementation wrongly
+required a list. Dedicated37466840210/1/job112279898907 therefore failed
+at that inventory check BEFORE any genuine checker invocation, after its
+complete focused suites passed74.486s/73.334s and source preflight succeeded.
+This is an author implementation/fixture error, not runner failure or a false
+scientific result. Original failed execution is retained in370/6016816642.
+
+The positive fixture was corrected FIRST to the native object. Its valid-case
+test then produced exactly one intended assertion failure and zero errors in
+each mode against77b1. Only afterwards the matcher was corrected to validate
+the exact dictionary type, key order AND all control values. Missing/reordered
+mappings, duplicate-label lists and wrong value types remain rejecting controls.
+The original scientific manifest was not edited. The complete147-case local
+results above are the successor mapping-based tests, not the older list-fixture
+results. The same147 keys and nine methods are retained, with four fixture
+construction lines corrected. All earlier list-fixture executions and original
+failed CI remain separate historical evidence; none is overwritten as success.
+
+The correction changes only the inventory predicate/fixed native mapping, the
+fixture's native inventory construction and this scope record. Full failure
+carriers, source preflight, postflight and all scientific bytes stay unchanged.
+An original review of77b1 cannot be represented as final successor review. The
+corrected source requires its own applicable fresh CI and exact-source review.
+
+## Review and integration boundary
+
+This is engineering validation, not the proof of FL, numerical certification,
+a false theorem finding, hostile-source sandbox, bounded-output service, durable
+storage guarantee or independent formal alignment. Existing mathematical scope
+and status remain. No complete local repository checkout or local Lean execution
+is claimed. Original failures and observation-only records remain historical;
+old CI cannot be rebound to the successor repair or a later main.
+
+Require an actual nonauthor engineering review of this final three-file delta,
+all applicable fresh dedicated/general checks on its actual candidate, and a
+separate eligible nonauthor/nonreader integrator's live ownership/ref/diff/source/
+review/thread reconciliation before ordinary expected-head merge and landed/push
+verification. #322 remains open for the broader wrapper class. No new owner
+approval, duplicate scientific register, main-tip lock or timer is requested.