d6g8k5htny-coder · Oct 6, 2026 · Oct 5, 2026 · Oct 5, 2026 · Oct 6, 2026 · Oct 6, 2026
@@ -0,0 +1,65 @@
+name: D3 L4 anisotropic coefficient replay
+
+'on':
+  pull_request:
+    paths:
+      - 'coefficients/d3_l4_anisotropic_20261005/**'
+      - 'tools/d3_l4_replay.py'
+      - 'tests/test_d3_l4_replay.py'
+      - '.github/workflows/d3-l4-anisotropic.yml'
+      - 'reviews/iba1_periodic_jet_claude_20261005/periodic_jet_check.py'
+      - 'coefficients/side24_v1/PROOF.md'
+      - 'coefficients/side24_v1/ENCLOSURE.json'
+      - 'imports/lifetime_parent_20260925/UNIFORM_MATRIX_CAP_AND_LIFETIME.md'
+  push:
+    branches: [main]
+    paths:
+      - 'coefficients/d3_l4_anisotropic_20261005/**'
+      - 'tools/d3_l4_replay.py'
+      - 'tests/test_d3_l4_replay.py'
+      - '.github/workflows/d3-l4-anisotropic.yml'
+      - 'reviews/iba1_periodic_jet_claude_20261005/periodic_jet_check.py'
+      - 'coefficients/side24_v1/PROOF.md'
+      - 'coefficients/side24_v1/ENCLOSURE.json'
+      - 'imports/lifetime_parent_20260925/UNIFORM_MATRIX_CAP_AND_LIFETIME.md'
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  d3-l4-replay:
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    strategy:
+      fail-fast: false
+      matrix:
+        mode: [normal, optimized]
+    name: d3-l4-replay (${{ matrix.mode }})
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262
+        with:
+          persist-credentials: false
+          fetch-depth: 0
+      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065
+        with:
+          python-version: '3.13.5'
+      - name: Test replay protocol and execute exact pinned numerical packet
+        shell: bash
+        run: |
+          set -euo pipefail
+          mkdir -p .d3-l4-protocol-evidence
+          FLAGS="-B -S"
+          if [ "${{ matrix.mode }}" = optimized ]; then FLAGS="-B -O -S"; fi
+          python $FLAGS -m unittest discover -s tests -v 2>&1 | tee ".d3-l4-protocol-evidence/${{ matrix.mode }}.log"
+          python -B -S tools/d3_l4_replay.py --mode "${{ matrix.mode }}" --output ".d3-l4-evidence/${{ matrix.mode }}"
+      - name: Preserve success or failure logs
+        if: always()
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
+        with:
+          name: d3-l4-${{ matrix.mode }}-${{ github.run_id }}-${{ github.run_attempt }}
+          path: |
+            .d3-l4-evidence/${{ matrix.mode }}/
+            .d3-l4-protocol-evidence/
+          include-hidden-files: true
+          if-no-files-found: error
@@ -0,0 +1,99 @@
+# D3 L4 replay coverage and execution custody
+
+Object: R2-D3-L4-EXECUTION-20261005-v1.
+Author: OpenAI / GPT-6 Astra Pro, session `d3-l4-round2-audit-20261005`;
+Dylan Roy — delegated AI work. Scientific effect NONE. Author will not self-merge.
+
+## Finding and frozen mathematical input
+
+AUD-D3L4-EXEC-001: the green general checks on Math-#304 did not execute the
+new coefficient calculation. Run 37383541867 had downstream, formal and aggregate
+jobs, but no dedicated `d3-l4` replay. Those successful checks are valid at their
+own scope, not numerical execution evidence for this packet.
+
+This is an isolated follow-on to Math-#304. Keep its head
+`5059e6d7049c5d65cd1e813545310bc71b86a12f` and all ten files under
+`coefficients/d3_l4_anisotropic_20261005/` unchanged while its nonauthor mathematical
+read is active. Claude's actual pickup is Math-#304 comment 6004751792; the
+acknowledgment and separate engineering scope are comment 6004801985.
+
+The frozen `SOURCE_FILES.json` has SHA-256
+`bba2ab2e351b644bbd3eec3b468ab88a2969242db9345f619294d4a8770a2f34`.
+It names nine packet payload files and four upstream pins. The driver authenticates
+that manifest before reading its instructions, verifies exact directory membership,
+all payload sizes/hashes/git blobs, all upstream git blobs, and the additional
+upstream SHA-256 when supplied. Symlinks and path aliases/traversals are rejected.
+This manifest is a custody root; neither a hash nor a replay is mathematical review.
+
+## Execution contract
+
+The dedicated workflow runs on relevant pull requests and main pushes, with path
+coverage for the entire packet, the replay machinery and every named upstream
+source. It also allows manual diagnostics. It has read-only contents permission,
+no persisted checkout credentials, full-commit action pins and Python 3.13.5.
+It does not edit a ruleset, change the general aggregate, or convert a manual run
+into required-PR evidence.
+
+Each of the normal and optimized matrix jobs runs the full repository `tests/`
+unittest discovery. It then executes the original seven packet tests, full N=128
+coefficient calculation, N=4 nonperiodic reference calculation, and source crosscheck.
+The main coefficient and reference outputs must match their frozen files byte for
+byte, with exit zero and empty stderr. A test process must actually report seven
+successful tests; zero collected tests do not count as a pass.
+
+For crosschecks, all interval strings, gamma strings, booleans, keys and other
+fields must match exactly. Only the five fields explicitly named
+`independent_disk_midpoint_diagnostic` may vary by at most 1e-10 across libm builds.
+These are already labeled noncertificate floating diagnostics in the source;
+this portability allowance is NOT added to or substituted for any certificate
+error. Nonfinite values, including exponent overflow such as `1e999`, are rejected.
+Duplicate JSON keys, malformed payloads and wrong JSON types are rejected.
+
+The execution driver, regression file and workflow are checked against the actual
+Git HEAD before and after replay. The driver must execute from that checkout.
+On GitHub the recorded HEAD must equal GITHUB_SHA, and repository/run/attempt
+identity must be present and consistent. A dirty execution tool cannot claim a
+clean tested commit. Source pins are verified again after execution.
+
+Commands have bounded timeouts. A crash, wrong exit, unexpected stderr, changed
+result or timeout is failure, never a successful negative control. Raw stdout,
+stderr and process metadata are retained, including a failed receipt on failures
+after evidence-directory creation. Existing evidence directories are not reused
+or overwritten. Workflow logs/artifacts survive failures through `if: always()`.
+The receipt always leaves scientific acceptance and independent review false.
+
+## Fresh local evidence and limits
+
+Twenty replay-protocol test methods pass under both `-B -S` and `-B -O -S`.
+They include real Python processes for success, crash, wrong exit, stderr,
+timeout and CLI-failure receipts, plus a temporary real Git repository to test
+dirty execution-source rejection. The source fixtures are explicitly synthetic,
+not a substitute for the mathematical packet.
+
+The initial 17-method suite failed because the driver was absent. A subsequent
+JSON exponent-overflow test and execution-source binding test each failed before
+their guards were implemented. An intermediate test-only module-reload exception-
+identity error was corrected by caching the imported test module; no numerical
+source was changed. These local development failures are retained in the audit
+bundle rather than represented as historical successes.
+
+The original N=128 numerical calculation and the source crosscheck were also
+freshly run in both modes in this round; their outputs match the delivered files
+exactly. This is author-side execution, not the provider-distinct mathematical
+review requested on #304.
+
+A complete local repository checkout could not be downloaded in this environment;
+therefore complete local repository tests and an end-to-end local run of the new
+CLI with all four upstream payloads are NOT claimed. The hosted jobs are designed
+to supply that missing execution evidence. Their actual conclusions and tested
+commits must be read before integration; configured coverage is not a passed run.
+
+## Review and integration
+
+Review this engineering delta separately from the frozen coefficient proof.
+After #304 lands, the follow-on owner must reconcile the base, verify that all ten
+packet identities and four upstream pins are preserved, obtain fresh applicable
+hosted checks and the scoped nonauthor disposition, then hand off to an eligible
+integrator. Do not merge this proposal into the live mathematical-review branch.
+No audit verdict, theorem flag, original proof, parent workflow or scientific
+status register is changed by this proposal.