Add active evidence loop, KGE proposals and predictive CSI research harness

.github/workflows/csi-encoder-training.yml

@@ -5,6 +5,12 @@ on:
     paths:
       - 'scripts/csi_training/**'
       - 'scripts/field-context-probe/**'
+      - 'scripts/active-kge/**'
+      - 'scripts/pretrain-csi-predictive.py'
+      - 'scripts/pretrain-csi-transfer.py'
+      - 'scripts/benchmark-csi-intake.py'
+      - 'scripts/evaluate-active-evidence.py'
+      - 'scripts/verify-active-replay.py'
       - 'scripts/train-csi-encoder.py'
       - 'scripts/train-espfi.py'
       - 'scripts/improve-espfi.py'
@@ -13,6 +19,7 @@ on:
       - 'scripts/requirements-espfi-training.txt'
       - 'scripts/requirements-espfi-inference.txt'
       - 'scripts/requirements-csi-training.txt'
+      - 'scripts/requirements-zenodo-intake.txt'
       - 'scripts/tests/test_csi_training*.py'
       - 'scripts/tests/test_espfi*.py'
       - '.github/workflows/csi-encoder-training.yml'
@@ -37,6 +44,7 @@ jobs:
           python -m pip install torch==2.11.0 --index-url https://download.pytorch.org/whl/cpu
           python -m pip install -r scripts/requirements-csi-training.txt
           python -m pip install -r scripts/requirements-espfi-training.txt
+          python -m pip install -r scripts/requirements-zenodo-intake.txt
       - name: Validate real training code using temporary synthetic data
         env:
           PYTHONPATH: scripts
@@ -76,3 +84,30 @@ jobs:
         run: cargo +1.90.0 test --locked
       - name: Exercise synthetic evidence pipeline
         run: cargo +1.90.0 run --locked --quiet --bin ruview-field-context-probe
+      - name: Exercise active evidence lifecycle
+        run: cargo +1.90.0 run --locked --quiet --bin ruview-field-context-probe -- --active-loop
+
+  active-kge-contract:
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1
+        with:
+          persist-credentials: false
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020
+        with:
+          node-version: '22'
+      - name: Install exact KGE experiment dependency
+        working-directory: scripts/active-kge
+        run: npm ci --ignore-scripts --no-audit --no-fund
+      - name: Validate actual trained KGE and bounded artifact contract
+        working-directory: scripts/active-kge
+        run: npm test
+      - name: Install fixed Rust toolchain
+        run: rustup toolchain install 1.90.0 --profile minimal
+      - name: Feed actual KGE proposals through Rust authority boundary
+        run: |
+          node scripts/active-kge/cli.mjs --train "$RUNNER_TEMP/kge" --context context_1_8 > "$RUNNER_TEMP/ranking.json"
+          cargo +1.90.0 run --locked --quiet --manifest-path scripts/field-context-probe/Cargo.toml -- --active-loop > "$RUNNER_TEMP/reference.json"
+          cargo +1.90.0 run --locked --quiet --manifest-path scripts/field-context-probe/Cargo.toml -- --active-loop --ranking "$RUNNER_TEMP/ranking.json" > "$RUNNER_TEMP/candidate.json"
+          python scripts/verify-active-replay.py "$RUNNER_TEMP/reference.json" "$RUNNER_TEMP/candidate.json"

.github/workflows/rust-cross-target.yml

@@ -0,0 +1,73 @@
+name: Rust cross-target dependency check
+
+on:
+  push:
+    branches: [main, develop, 'feature/*', 'feat/*', 'hotfix/*']
+    paths:
+      - 'v2/Cargo.lock'
+      - 'v2/Cargo.toml'
+      - 'v2/rust-toolchain.toml'
+      - 'v2/crates/wifi-densepose-hardware/**'
+      - '.github/workflows/rust-cross-target.yml'
+  pull_request:
+    branches: [main, develop]
+    paths:
+      - 'v2/Cargo.lock'
+      - 'v2/Cargo.toml'
+      - 'v2/rust-toolchain.toml'
+      - 'v2/crates/wifi-densepose-hardware/**'
+      - '.github/workflows/rust-cross-target.yml'
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  dependency-path:
+    name: ${{ matrix.name }}
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 20
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - name: Windows MSVC dependency path
+            os: windows-2025
+            target: x86_64-pc-windows-msvc
+          - name: Android ARM64 dependency path
+            os: ubuntu-24.04
+            target: aarch64-linux-android
+    steps:
+      - name: Checkout code and path dependencies
+        uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262
+        with:
+          persist-credentials: false
+          submodules: recursive
+
+      - name: Install fixed Rust toolchain and target
+        uses: dtolnay/rust-toolchain@4cda84d5c5c54efe2404f9d843567869ab1699d4
+        with:
+          toolchain: 1.89.0
+          targets: ${{ matrix.target }}
+
+      - name: Configure the preinstalled Android NDK linker
+        if: matrix.target == 'aarch64-linux-android'
+        shell: bash
+        run: |
+          ndk_bin="$ANDROID_NDK_HOME/toolchains/llvm/prebuilt/linux-x86_64/bin"
+          linker="$ndk_bin/aarch64-linux-android24-clang"
+          archiver="$ndk_bin/llvm-ar"
+          ranlib="$ndk_bin/llvm-ranlib"
+          test -x "$linker"
+          test -x "$archiver"
+          test -x "$ranlib"
+          echo "$ndk_bin" >> "$GITHUB_PATH"
+          echo "CC_aarch64_linux_android=$linker" >> "$GITHUB_ENV"
+          echo "AR_aarch64_linux_android=$archiver" >> "$GITHUB_ENV"
+          echo "RANLIB_aarch64_linux_android=$ranlib" >> "$GITHUB_ENV"
+          echo "CARGO_TARGET_AARCH64_LINUX_ANDROID_LINKER=$linker" >> "$GITHUB_ENV"
+          echo "CARGO_TARGET_AARCH64_LINUX_ANDROID_AR=$archiver" >> "$GITHUB_ENV"
+
+      - name: Check hardware transport dependency path
+        working-directory: v2
+        run: cargo +1.89.0 check --locked -p wifi-densepose-hardware --lib --target ${{ matrix.target }}

docs/active-evidence-experiment.md

@@ -0,0 +1,176 @@
+# Active evidence, KGE proposals and predictive CSI experiment
+
+Research implementation, measured on October 13, 2026 UTC. No field accuracy or
+production readiness claim. Keep the existing sensing runtime and model defaults.
+
+## Implemented outcome
+
+The actual pinned RuField and WorldGraph libraries now run an authored sequence
+that transitions through empty, occupied, unknown, occupied and unknown. A
+conflicting observation creates uncertainty. Expired observations cannot keep an
+occupancy belief alive. Historical graph nodes remain provenance, not a current
+occupancy API. Repeated, derived, invalid, stale or future evidence is rejected.
+
+The separately trained actual `@ruvector/kge` package emits bounded ranking JSON.
+Rust accepts only three registered logical sensor names and finite scores. It
+reserves an authored packet, airtime and energy budget for proposals, never
+actuates. An integrated verifier compares the default and KGE replays and requires
+identical evidence and belief fields. Only the proposal can differ.
+
+The logical radar and door sources currently use signed WiFi feature fixtures to
+exercise the pinned rule. They are not implemented physical radar or door
+drivers. All replay observations, costs, relevance labels and thresholds are
+SYNTHETIC. KGE scores and fusion heuristics are not calibrated probabilities.
+
+## Decisions
+
+* [ADR 421: evidence lifecycle](adr/ADR-421-active-evidence-loop.md)
+* [ADR 422: KGE proposal authority](adr/ADR-422-kge-measurement-proposals.md)
+* [ADR 423: predictive representation experiment](adr/ADR-423-predictive-csi-representation-experiment.md)
+
+## Measured results
+
+| Experiment | Result | Decision |
+| --- | --- | --- |
+| Rust active loop, three rounds of 1,000 alternating pairs | Median 0.0771 to 0.0853 ms; p95 0.1187 to 0.1377 ms | Retain bounded research implementation |
+| Added Rust cost over signed single root fusion | Paired extra p95 0.0412 to 0.0576 ms | Small measured software overhead, no sensing benefit claim |
+| Actual KGE, six authored held out target links | Hits@1 6/6; frequency and shuffled graph controls 2/6 each | Retain synthetic integration test, not real utility evidence |
+| KGE reference ranking, three rounds of 1,000 calls | Median 0.0345 to 0.0349 ms; p95 0.0506 to 0.0577 ms | Retain original default |
+| Cached KGE validation experiment | Fails lower median and p95 in every round gate | Reject optimization; preserve initial failure receipt |
+| Predictive CSI, fixed five epochs on original training data only | Latent MSE 0.9740 versus global constant 0.9413 | Reject utility/promotion claim |
+| Predictive frozen encoder timing | Median 0.919 to 1.175 ms; p95 1.326 to 1.478 ms | Timing only; not a sensing model improvement |
+
+Rust and Python ran on the existing ruOS Linux x86_64 AMD EPYC environment.
+Rust 1.90.0 used the release profile and exact existing lockfile. Python used
+Torch 2.11.0+cpu and NumPy 2.4.4 with one compute thread. Node KGE ran separately
+on local Linux x64 Node 24.19.0, native backend. Do not add these timings to claim
+a measured end to end latency across machines. Rust excludes signing, acquisition,
+ranking transport and hardware. KGE excludes training and loading. Encoder timing
+includes input validation, log normalization, token encoding and pooling but
+excludes acquisition and file loading. This is not ESP32 timing or peak RAM.
+
+The predictive experiment fit 280 recordings and evaluated 139 inner development
+recordings from the original training partition. The remaining 418 training
+recordings were unused. No original validation or final test arrays were decoded.
+These development groups have been reused in previous research and are not a new
+independent confirmation set. Training took 7.34 seconds in the recorded scope.
+The same final teacher targets and masks were used for every loss comparator:
+untrained 1.1399, global constant 0.9413, per position constant 1.0487, trained
+0.9740. Effective recording rank was 1.4055 and triggered the collapse warning.
+Reloaded embeddings were bit identical. A loss improvement against an untrained
+network is insufficient when a constant predictor performs better.
+
+No candidate from this experiment replaces the activity classifier, the published
+eight feature encoder or any production model. No paid compute or hardware
+operation was performed. Existing ruOS CPU resources were used.
+
+## Reproduce the software acceptance test
+
+From the repository root, with Node, Python and Rust 1.90.0 installed:
+
+```sh
+cd scripts/active-kge
+npm ci --ignore-scripts --no-audit --no-fund
+npm test
+cd ../..
+cargo +1.90.0 test --locked --manifest-path scripts/field-context-probe/Cargo.toml
+node scripts/active-kge/cli.mjs --train /tmp/ruview-kge-new --context context_1_8 > /tmp/ruview-ranking.json
+cargo +1.90.0 run --release --locked --manifest-path scripts/field-context-probe/Cargo.toml -- --active-loop > /tmp/ruview-reference.json
+cargo +1.90.0 run --release --locked --manifest-path scripts/field-context-probe/Cargo.toml -- --active-loop --ranking /tmp/ruview-ranking.json > /tmp/ruview-candidate.json
+python scripts/verify-active-replay.py /tmp/ruview-reference.json /tmp/ruview-candidate.json
+```
+
+Use a new output directory for each KGE training run. The Linux file reader
+supports x86_64 and AArch64; other platforms must explicitly supply ranking JSON
+on stdin using `--ranking -`. Built replay executables perform no network or
+hardware activity. Initial npm/Cargo installation requires package access.
+
+Python training requires `scripts/requirements-csi-training.txt`. Run:
+
+```sh
+PYTHONPATH=scripts OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 python -m unittest discover -s scripts/tests -p 'test_csi_training*.py' -q
+python scripts/pretrain-csi-predictive.py --synthetic --output /tmp/ruview-predictive-new --epochs 5 --max-seconds 120
+```
+
+For the public run, replace `--synthetic` with `--manifest` pointing to the
+existing ADR 420 verified prepared manifest, with five sample temporal pooling.
+Do not reinterpret different input axes or change splits to obtain a favorable
+result. See [predictive harness contract](../scripts/csi_training/PREDICTIVE.md).
+
+## Validation and security
+
+The completed focused suites passed 41 Python tests, 14 Rust tests and 5 KGE
+tests. The integrated replay verifier passed. An independent review found no
+remaining source blocker after fixes. The first remote Python run lacked the
+unchanged collector helper; restoring that test dependency produced the passing
+run. A KGE test initially depended on the current directory; its CLI path now
+resolves relative to the test file and passes from the repository root.
+
+| Boundary | Check and residual limit |
+| --- | --- |
+| Observation admission | Signature, unique root/hash, monotonic source and application clocks, explicit expiry, derived root rejection; public fixture keys are not production enrollment |
+| Rejected observation | Does not advance the accepted clock or mutate the graph; exhausted capacity intentionally sets a sticky safety state |
+| Ranking file | Open once with nofollow and nonblocking flags, validate the opened regular descriptor, bound bytes, reject duplicate JSON keys, unknown fields and nonfinite scores |
+| KGE artifact | Bound file reads and native dimensions; exact authored vocabulary, relations, indexed triples, training splits and training receipt binding; checksums are not signer authentication |
+| Predictive artifact | Bound descriptor reads, strict JSON, exact finite tensor shape/dtype checks, safetensors, source hash before/after and reload parity; local artifact trust still required |
+| Authority | No hardware, network endpoint or executable supplied by ranking; proposals cannot modify roots or promote models |
+| Evaluation | Unknown counts as an occupied miss; separate unknown coverage limits prevent abstention from manufacturing gains; bootstrap by room |
+
+Scoped npm audit of the exact KGE lockfile returned zero reported vulnerabilities
+on October 13, 2026 UTC. This is not a security audit of all RuView dependencies,
+the native binary internals or external device firmware. Rust pins and its lockfile
+were unchanged. No fresh Rust advisory database audit was performed.
+
+## Field measurement contract
+
+`python scripts/evaluate-active-evidence.py /path/to/aligned-intervals.json`
+computes aggregate duration metrics and seeded room clustered percentile
+intervals. It accepts schema version 1, declared evidence, and sessions with
+room/session IDs, matching baseline/candidate packets, airtime, energy and sensor
+availability, and contiguous aligned intervals with independently supplied empty
+or occupied truth plus empty/occupied/unknown predictions. Input is a bounded
+regular JSON file. See `test_csi_training_active_evaluation.py` for the minimal
+synthetic input generator.
+
+The default CLI remains report-only for interactive analysis. Automation can add
+`--require-pass`: a valid report still prints in full, then exits 3 when
+`overallGate` is `fail` or `inconclusive`; a passing report exits 0 and malformed
+input exits 2. This process status enforces only the frozen software gate. It does
+not change `fieldAcceptance`, authorize promotion or authenticate supplied labels.
+
+The numerical gate requires at least 20% lower false occupied duration, a 95%
+room clustered interval supporting reduction, recall delta lower bound at least
+minus one percentage point, and unknown duration delta upper bound at most one
+percentage point separately for empty and occupied intervals. Zero baseline
+false alarms or inadequate bootstrap exposure produce inconclusive results.
+Two rooms are a computational minimum, not a claim of sufficient statistical
+power. Pilot data must size a separate confirmation study.
+
+Labels, actual costs, fresh splits and capture completeness require independent
+verification. Consequently even a numerical pass always returns
+`fieldAcceptance: not_established` and `promotionAuthorized: false`. The paired
+transition gate requires empty-to-occupied and occupied-to-empty events in at
+least two rooms. Room-clustered upper confidence bounds must show no increase in
+interval-aligned delay, unresolved transitions or false-occupied episodes per
+empty hour. Unknown predictions extend delay; a transition that never resolves
+within its truth run is retained with that run's full duration. `overallGate`
+passes only when both duration and transition gates pass. Stationary occupant
+sensitivity and sub-interval acquisition latency still require field analysis.
+Replay cannot compare an action for which no eligible measurement was captured.
+
+## Receipts and rollback
+
+Receipts contain aggregate results and hashes, not captures or learned tensors:
+
+* `validation/active-loop-benchmark-20261013.json`
+* `validation/active-kge-field-replay-20261013.json`
+* `validation/active-kge-20261013.json`
+* `validation/active-kge-cache-rejected-20261013.json`
+* `validation/active-predictive-public-20261013.json`
+* `validation/active-evidence-verification-20261013.json`
+
+Rollback is removal of the isolated experiment and its CI additions. Existing
+production paths and published weights are unchanged. The next useful investment
+is independently annotated continuous capture, followed by frozen room/session
+confirmation and an inductive sensor utility experiment, not promotion of the
+rejected predictive checkpoint.