Integrate native OpenCode CPU path and bounded coding trial

.github/workflows/opencode-inference.yml

@@ -15,6 +15,14 @@ on:
         options:
           - qwen3-0.6b-v1
           - qwen3-4b-instruct-2507-v1
+      inference_backend:
+        description: Explicit owner-selected backend; native CPU requires the reviewed 4B profile
+        type: choice
+        required: true
+        default: torch
+        options:
+          - torch
+          - llama_cpp_bf16_v1
 
 permissions:
   contents: read
@@ -30,6 +38,7 @@ jobs:
     env:
       EXPECTED_CODE_SHA: ${{ inputs.expected_code_sha }}
       MODEL_PROFILE: ${{ inputs.model_profile }}
+      INFERENCE_BACKEND: ${{ inputs.inference_backend }}
       PYTHONDONTWRITEBYTECODE: '1'
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
@@ -41,7 +50,7 @@ jobs:
         id: selected_core
         run: |
           set -euo pipefail
-          python3 -B scripts/opencode_ci.py select --model-profile "$MODEL_PROFILE" >>"$GITHUB_OUTPUT"
+          python3 -B scripts/opencode_ci.py select --model-profile "$MODEL_PROFILE" --inference-backend "$INFERENCE_BACKEND" >>"$GITHUB_OUTPUT"
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           repository: VOLPAROSSA/volparossa
@@ -52,7 +61,7 @@ jobs:
         run: |
           set -euo pipefail
           python3 -B scripts/opencode_ci.py source --core "$PWD/build/ci-core" \
-            --expected-code "$EXPECTED_CODE_SHA" --model-profile "$MODEL_PROFILE"
+            --expected-code "$EXPECTED_CODE_SHA" --model-profile "$MODEL_PROFILE" --inference-backend "$INFERENCE_BACKEND"
       - name: Install official tools only on the disposable GitHub host
         run: |
           set -euo pipefail
@@ -81,14 +90,14 @@ jobs:
           export XDG_RUNTIME_DIR
           export DBUS_SESSION_BUS_ADDRESS="unix:path=$XDG_RUNTIME_DIR/bus"
           printf 'XDG_RUNTIME_DIR=%s\nDBUS_SESSION_BUS_ADDRESS=%s\n' "$XDG_RUNTIME_DIR" "$DBUS_SESSION_BUS_ADDRESS" >>"$GITHUB_ENV"
-          python3 -B scripts/opencode_ci.py preflight --model-profile "$MODEL_PROFILE"
+          python3 -B scripts/opencode_ci.py preflight --model-profile "$MODEL_PROFILE" --inference-backend "$INFERENCE_BACKEND"
       - name: Source-build the pinned OpenCode runtime (no model)
         timeout-minutes: 55
         run: bash scripts/opencode_ci_build.sh
       - name: Fetch exact public Node and Debian guest image
         run: |
           set -euo pipefail
-          python3 -B scripts/opencode_ci.py assets --core "$PWD/build/ci-core" --model-profile "$MODEL_PROFILE"
+          python3 -B scripts/opencode_ci.py assets --core "$PWD/build/ci-core" --model-profile "$MODEL_PROFILE" --inference-backend "$INFERENCE_BACKEND"
           image_url=$(jq -er '.url' build/ci-core/tests/helper/debian13-amd64-image-v1.json)
           image="$PWD/build/debian-13-genericcloud-amd64-20260826-2582.qcow2"
           curl --fail --silent --show-error --location --connect-timeout 30 --max-time 1200 \
@@ -102,8 +111,8 @@ jobs:
           test "$(git rev-parse HEAD)" = "$EXPECTED_CODE_SHA"
           test -z "$(git status --porcelain)"
           python3 -B scripts/smoke_opencode_inference.py pack --core "$PWD/build/ci-core" \
-            --node "$PWD/build/ci-node/bin/node" --output "$PWD/build/ci-inputs.tar.gz" --model-profile "$MODEL_PROFILE"
-          python3 -B scripts/opencode_ci.py capture --model-profile "$MODEL_PROFILE"
+            --node "$PWD/build/ci-node/bin/node" --output "$PWD/build/ci-inputs.tar.gz" --model-profile "$MODEL_PROFILE" --inference-backend "$INFERENCE_BACKEND"
+          python3 -B scripts/opencode_ci.py capture --model-profile "$MODEL_PROFILE" --inference-backend "$INFERENCE_BACKEND"
       - name: One actual OpenCode core Qwen edit and test trial
         timeout-minutes: 115
         run: |
@@ -115,6 +124,7 @@ jobs:
             XDG_RUNTIME_DIR="$XDG_RUNTIME_DIR" DBUS_SESSION_BUS_ADDRESS="$DBUS_SESSION_BUS_ADDRESS" \
             python3 -B scripts/smoke_opencode_inference.py execute --yes --host-tools-profile github-ubuntu-24.04 \
               --model-profile "$MODEL_PROFILE" \
+              --inference-backend "$INFERENCE_BACKEND" \
               --core "$PWD/build/ci-core" --tools /usr \
               --image "$PWD/build/debian-13-genericcloud-amd64-20260826-2582.qcow2" \
               --bundle "$PWD/build/ci-inputs.tar.gz" --output "$PWD/build/ci-vm"

.github/workflows/source-checks.yml

@@ -4,7 +4,6 @@ on:
   push:
     branches: [main]
   pull_request:
-    branches: [main]
 
 permissions:
   contents: read
@@ -27,6 +26,8 @@ jobs:
         run: python3 -B -m unittest discover -s tests -p 'test_build_codex_runtime.py'
       - name: Check OpenCode source-build and namespace contracts
         run: python3 -B -m unittest discover -s tests -p 'test_*opencode*.py'
+      - name: Check explicit output-constraint session selection
+        run: python3 -B -m unittest discover -s tests -p test_output_constraint_session.py
 
 # No dependency installation, runtime/model download, real inference, editor
 # installation or privileged service startup. These are source/protocol checks.

README.md

@@ -63,16 +63,31 @@ not Codex. It joins these implemented components:
 - An explicit Linux launcher with a selected writable project, temporary state,
   no inherited account credentials and no external network interface.
 
-**Current proof:** the pinned OpenCode source builds and the actual runtime
+**Runtime and adapter proof:** the pinned OpenCode source builds and the actual runtime
 completes a tool loop through the production launcher and adapters: one approved
 command changes a disposable file, its tool result returns to the core interface,
 and the session shuts down cleanly. A second native trial invokes the cooperative
 tool, preserves complete and incomplete core results, and confirms that only the
 enrolled public snapshot crosses the bridge. Both trials use **synthetic core/model
 replies**, not real inference or peer execution. Focused checks additionally cover
-adapter, editor and lifecycle behavior. The complete model-driven OpenCode
-read/edit/test loop, native editor UI operation, protected peer execution and a
-finished immune-policy path remain unproved.
+adapter, editor and lifecycle behavior.
+
+**Real model-driven coding:** the [Qwen4B native CPU trial](https://github.com/VOLPAROSSA/volparossa-code/actions/runs/38824828727)
+passes the bounded coding task **and its complete lifecycle**. Actual OpenCode
+and core inference produce two reads, an approved edit, an approved test and
+a final response. The original tests stay unchanged and an independent check
+passes. Task-resource measurements, unchanged inputs, clean core shutdown and
+guest/VM cleanup are verified.
+
+The task takes about 37 minutes on the disposable CPU runner; this is a working
+development slice, not competitive performance or general coding-quality proof.
+The core processes the complete prompt in adaptive blocks while retaining the
+model, permissions, resource budgets and ten-second owner-control guard. Earlier
+failed trials remain documented. This path pins the exact tested core candidate
+`7d0f9f0b`, rather than selecting whichever core happens to be on `main`.
+Native editor UI operation, protected peer execution and the complete
+immune-policy path remain open. See the
+[native CPU evidence](docs/OPENCODE.md#explicit-native-cpu-candidate).
 
 **Real public peer proof:** a separate owner command successfully obtained a
 small single-file replacement from a real Qwen0.6B worker on another peer.
@@ -138,9 +153,8 @@ notice and source pin are preserved in [third-party provenance](THIRD_PARTY_LICE
 
 ## Remaining work
 
-- Exercise the source-built OpenCode runtime with actual core inference and a
-  model-driven read/edit/test task with explicit local approvals and independently
-  checked results, then verify native editor operation.
+- Verify native editor operation with the proven real-model read/edit/test
+  lifecycle, and improve measured coding quality, speed and resource use.
 - Extend the proven public single-file owner command to a complete real-model
   OpenCode cooperative tool loop; retain original results, cancellation and provenance.
 - Implement remote conversation execution and actual protected private work,