CaPU × ATMAN: loopback HTTP recovery with a separate read-only observer

.github/workflows/capu-atman-http-recovery-v2.yml

@@ -0,0 +1,35 @@
+name: CaPU ATMAN loopback HTTP recovery
+on:
+  pull_request:
+    paths:
+      - 'experiments/capu_atman_recovery_v1/**'
+      - 'experiments/capu_atman_http_recovery_v2/**'
+      - '.github/workflows/capu-atman-http-recovery-v2.yml'
+  workflow_dispatch:
+permissions:
+  contents: read
+jobs:
+  bounded-http:
+    runs-on: ubuntu-latest
+    timeout-minutes: 8
+    strategy:
+      fail-fast: false
+      matrix:
+        python: ['3.11', '3.13']
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python }}
+      - run: python -m pip install -r experiments/capu_atman_recovery_v1/requirements.txt
+      - run: python experiments/capu_atman_recovery_v1/bootstrap.py
+      - run: python experiments/capu_atman_recovery_v1/run.py --output evidence-v1
+      - run: python experiments/capu_atman_http_recovery_v2/run.py --output evidence-v2
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: capu-atman-http-python-${{ matrix.python }}
+          path: |
+            evidence-v1/
+            evidence-v2/
+          if-no-files-found: error

.github/workflows/capu-atman-recovery-v1.yml

@@ -0,0 +1,34 @@
+name: CaPU ATMAN recovery laboratory
+on:
+  pull_request:
+    paths:
+      - 'experiments/capu_atman_recovery_v1/**'
+      - '.github/workflows/capu-atman-recovery-v1.yml'
+  workflow_dispatch:
+permissions:
+  contents: read
+jobs:
+  bounded-composition:
+    runs-on: ubuntu-latest
+    timeout-minutes: 8
+    strategy:
+      fail-fast: false
+      matrix:
+        python: ['3.11', '3.13']
+    defaults:
+      run:
+        working-directory: experiments/capu_atman_recovery_v1
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python }}
+      - run: python -m pip install -r requirements.txt
+      - run: python bootstrap.py
+      - run: python run.py --output evidence
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: capu-atman-recovery-python-${{ matrix.python }}
+          path: experiments/capu_atman_recovery_v1/evidence/
+          if-no-files-found: error

experiments/capu_atman_http_recovery_v2/.gitignore

@@ -0,0 +1,2 @@
+__pycache__/
+evidence/

experiments/capu_atman_http_recovery_v2/README.md

@@ -0,0 +1,92 @@
+# CaPU × ATMAN — bounded loopback HTTP recovery v2
+
+This experiment extends the **process boundary**, not the production claim.
+The v1 source and PR #103 head (`977864167c65f161e6db87b3d14257a11a67516f`)
+remain unchanged. This candidate belongs on a separate branch and **draft PR**.
+No merge, deployment, main publication, or ready-for-review transition.
+
+## What is actually external?
+
+1. A controller subprocess owns only its controller SQLite path and a loopback URL.
+2. A separate HTTP device process executes a non-idempotent counter append in its
+   own SQLite ledger. The durable insertion **is** the effect in this lab.
+3. A third, stdlib-only observer process opens that device ledger read-only. It
+   counts actual requests/effects and hashes the observed rows, without importing
+   controller code or reading controller state.
+
+This is a real TCP/HTTP and process boundary on **one trusted host**, not a real
+payment, physical actuator, third-party deployment, WAN or independent attestor.
+Process separation does not create OS permission separation. The device deliberately
+accepts direct lab calls; bypass-resistant endpoint enforcement is out of scope.
+
+## Central scenario
+
+```
+ATMAN authority -> durable A6 UNKNOWN reservation -> one HTTP POST
+ -> device effect committed -> response held/lost
+ -> independent observer sees effect -> controller process killed
+ -> new controller process refuses retry while UNKNOWN
+ -> policy changes -> exact historical positive receipt arrives
+ -> A7 reconciliation commits outcome -> new attempts remain blocked
+```
+
+The device database path is never passed to a controller worker. The observer
+triggers the kill only after it has read the effect, while the controller is
+still awaiting the held response. The controller and device have no shared
+transaction. A receipt query with no matching row returns UNKNOWN, never a
+manufactured NOT_COMMITTED. HTTP 503, malformed JSON and socket timeout do not
+release the reservation. The HTTP client makes one request with no automatic
+retry, redirect or proxy behavior.
+
+## Tests and equal-guarantee comparison
+
+There are 14 named scenarios for each of two arms (28 tests): normal ACK,
+connection loss after/before the effect, HTTP 503 after the effect, malformed ACK,
+observer-triggered controller kill, timeout, device-process restart, receipt
+tampering, full-token receipt lookup, competing dispatch workers, stale authority,
+receipt replay, and unguarded direct-call duplication.
+
+The ordinary FSM arm shares HTTP/SQLite I/O, ATMAN authority verification and the
+synthetic receipt primitive, but not native CaPU lifecycle transitions. The runner
+compares the recorded controller states and **independently observed device rows**
+for every named scenario, not merely test pass counts.
+
+The unguarded control intentionally sends two identical HTTP effect requests:
+they must produce two effects. An ambiguous duplicate ledger yields CONFLICT,
+not a fabricated clean COMMITTED receipt. This control also demonstrates the
+explicit lack of bypass resistance; it is not an architectural superiority claim.
+
+## Reproduce
+
+From repository root (Python 3.11 or 3.13):
+
+```sh
+python -m pip install -r experiments/capu_atman_recovery_v1/requirements.txt
+python experiments/capu_atman_recovery_v1/bootstrap.py
+python experiments/capu_atman_recovery_v1/run.py --output evidence-v1
+python experiments/capu_atman_http_recovery_v2/run.py --output evidence-v2
+```
+
+The isolated workflow bootstraps the three pinned upstream modules, runs both
+suites, and uploads results. `result.json` includes source hashes, environment,
+all observations and a digest. Check the exit code and actual CI run: a local
+PASS is not CI PASS or review approval. Source verification pins the reused v1
+bootstrap and adapter; upstream modules are verified by the unchanged bootstrap.
+
+## Non-claims / review boundary
+
+- Public deterministic fixture keys and A7 synthetic rotate/XOR receipt tags are
+  unchanged. No production authentication or cryptographic security claim.
+- One trusted controller/device/lineage, one unresolved attempt. No general
+  exactly-once or liveness guarantee; UNKNOWN may remain blocked indefinitely.
+- No physical power-loss/NVRAM, storage rollback, Byzantine device, malicious
+  host, distributed atomic commit or production transport proof.
+- Authorization linearizes at durable admission, not at physical actuation.
+  A late positive receipt records history; it does not renew old authority.
+- The observer is a separate observation path, not an independent organization,
+  timestamp, trust anchor, or proof that all external events were recorded.
+- No CPU/FPGA speed, energy, full ATMAN, Bardo/COSMIC integration, or advantage
+  over the equal-guarantee ordinary FSM is established.
+- Only native Codex review is the review gate. One request per final candidate
+  head; do not mutate that head after review. Keep draft and unmerged if quota
+  prevents review. Tests or manual inspection are not review approval.