Harden Stackchan controls and release authority

.gitattributes

@@ -1,3 +1,6 @@
 companion/gradlew text eol=lf
+tools/release_toolchain_identity_allowlist.json text eol=lf
+tools/release_toolchain_identity.ps1 text eol=lf
+tools/verify_git_pack_semantics.py text eol=lf
 media/voice/rvc/model.pth filter=lfs diff=lfs merge=lfs -text
 media/voice/rvc/model.index filter=lfs diff=lfs merge=lfs -text

.github/workflows/firmware.yml

@@ -83,16 +83,20 @@ jobs:
 
   bridge-tests:
     runs-on: ubuntu-latest
+    timeout-minutes: 15
     steps:
       - uses: actions/checkout@v7
         with:
           ref: ${{ env.STACKCHAN_CI_SOURCE_SHA }}
 
       - uses: actions/setup-python@v6
         with:
-          python-version: "3.12"
+          python-version: "3.12.10"
+          cache: pip
+          cache-dependency-path: bridge/requirements-vision.txt
 
       - name: Install bridge test dependencies
+        timeout-minutes: 5
         run: |
           sudo apt-get update
           sudo apt-get install -y ffmpeg
@@ -467,7 +471,7 @@ jobs:
       - uses: actions/setup-python@v6
         if: matrix.setup_desktop_runtime
         with:
-          python-version: "3.12"
+          python-version: "3.12.10"
 
       - uses: actions/setup-java@v5
         with:
@@ -502,7 +506,7 @@ jobs:
         run: |
           $runtimeRoot = Join-Path "${{ github.workspace }}" "output/desktop-python-runtime/${{ matrix.desktop_platform }}"
           ./tools/prepare_desktop_python_runtime.ps1 `
-            -SourcePython (Get-Command python).Source `
+            -SourcePython (Get-Command python -CommandType Application -ErrorAction Stop).Source `
             -RuntimeRoot $runtimeRoot `
             -SourceName "github-actions-pr-${{ runner.os }}-${{ runner.arch }}-python-3.12" `
             -Force `
@@ -694,14 +698,18 @@ jobs:
 
       - uses: actions/setup-python@v6
         with:
-          python-version: "3.12"
+          python-version: "3.12.10"
 
       - name: Install PlatformIO
-        run: python -m pip install --upgrade pip platformio
+        run: python -m pip install --upgrade pip -r requirements-firmware-release.txt
 
       - name: Run native logic tests
         run: pio test -e native_logic
 
+      - name: Verify dedicated wake-capture release boundary
+        shell: pwsh
+        run: ./tools/test_dedicated_wake_capture_contract.ps1
+
       - name: Compile native logic with Glow persona
         run: pio test -e native_logic --without-testing
         env:
@@ -716,7 +724,7 @@ jobs:
 
       - uses: actions/setup-python@v6
         with:
-          python-version: "3.12"
+          python-version: "3.12.10"
 
       - name: Verify Windows bridge launch contracts
         shell: pwsh
@@ -726,11 +734,29 @@ jobs:
           ./tools/test_stackchan_dashboard_launcher_contract.ps1
 
       - name: Install PlatformIO
-        run: python -m pip install --upgrade pip platformio
+        run: python -m pip install --upgrade pip -r requirements-firmware-release.txt
+
+      - name: Run firmware reproducibility contract
+        shell: pwsh
+        run: ./tools/test_firmware_reproducible_build_contract.ps1
 
       - name: Build firmware
         run: pio run -e stackchan -e stackchan_servo_calibration
 
+      - name: Build secret-free public full firmware
+        shell: pwsh
+        run: |
+          $env:PLATFORMIO_CORE_DIR = Join-Path $env:RUNNER_TEMP "stackchan-pioarduino"
+          pio run -e stackchan_release_full
+
+      - name: Probe installed firmware compilers for path normalization
+        shell: pwsh
+        run: |
+          ./tools/test_firmware_reproducible_build_contract.ps1
+          $pioarduinoCoreDir = Join-Path $env:RUNNER_TEMP "stackchan-pioarduino"
+          ./tools/test_firmware_reproducible_build_contract.ps1 `
+            -CompilerProbeCoreDir $pioarduinoCoreDir
+
       - name: Build unit-test firmware
         run: pio test -e stackchan --without-uploading --without-testing
 
@@ -743,3 +769,7 @@ jobs:
             .pio/build/stackchan/firmware.elf
             .pio/build/stackchan/partitions.bin
             .pio/build/stackchan/bootloader.bin
+            .pio/build/stackchan_release_full/firmware.bin
+            .pio/build/stackchan_release_full/firmware.elf
+            .pio/build/stackchan_release_full/partitions.bin
+            .pio/build/stackchan_release_full/bootloader.bin

AGENTS.md

@@ -64,10 +64,12 @@ Run the narrow tests for the code touched, then the relevant broad gates:
 pio test -e native_logic
 python -m unittest discover -s bridge -p "test_*.py"
 python bridge/trusted_facts_smoke.py --memory-file output/pc-brain/latest/memory.json --json
+powershell.exe -NoProfile -ExecutionPolicy Bypass -File tools\test_firmware_reproducible_build_contract.ps1
 pio run -e stackchan_release_full
 powershell.exe -NoProfile -ExecutionPolicy Bypass -File tools\test_full_system_soak_evidence_contract.ps1
 powershell.exe -NoProfile -ExecutionPolicy Bypass -File tools\test_current_lead_reproducibility_contract.ps1
 powershell.exe -NoProfile -ExecutionPolicy Bypass -File tools\test_archive_current_lead_contract.ps1
+powershell.exe -NoProfile -ExecutionPolicy Bypass -File tools\test_passive_no_motion_evidence_contract.ps1
 ```
 
 Use `tools/check_full_system_soak_evidence.ps1` for completed hardware runs. A build or short
@@ -82,6 +84,21 @@ robot failure; use repeated endpoint, process, bridge-socket, and runtime eviden
 `stackchan_release_full` is the secret-free public build. Per-device `stackchan_camera_probe` or
 `stackchan_release_forensics` builds require explicit private OTA/pairing configuration and must
 never be substituted into a public package or GitHub release asset.
+Firmware reproducibility is narrowly scoped to the same clean Git commit, operating system,
+dependency/toolchain bytes, canonical recorded PlatformIO configuration, and no listed ambient
+build overrides. The hook maps lexical/resolved project and core paths to stable prefixes. Every
+Arduino firmware environment must inherit exactly one `platformio_reproducible_build.py` pre-hook;
+`native_logic` must inherit none.
+Release packages must use two distinct short detached build roots, isolated empty per-cycle build
+caches, exact cycle-B dependency snapshots, and a separate clean commit-pinned source worktree for
+all tracked package inputs. Dependency evidence must select the exact verbose-resolved platform and
+only resolved shared-core packages. Failed logs are moved to private evidence while the complete
+failed detached worktree remains attached for inspection. The trusted checkout verifier must never
+execute package-contained code, repository-local Git hooks, or fsmonitor configuration.
+A `diagnostic-*` package is inspection-only: its firmware and dependency identity are unbound, it
+must not be flashed/published or used as evidence, and `-AllowDirtyPackage` never grants that authority.
+Matching rebuild hashes prove build determinism only. They do not transfer physical qualification,
+soak, stability, or safety evidence between different SHA-256 binaries.
 
 ## Change Discipline
 

ALIVENESS_THESIS.md

@@ -0,0 +1,105 @@
+# Aliveness Thesis
+
+Status: design hypothesis, not a capability or evidence claim
+Baseline: `39b750e6c354d1c4721c70bf20fba98b8ce5c3ec`
+Last revised: 2026-08-02
+
+## Thesis
+
+Stackchan should feel alive when its present behavior is a truthful, timely consequence of its
+recent interaction, bounded memory, current body state, permitted perception, commitments, and
+observed outcomes. The effect must come from causal coherence across time, not from asserting
+humanity, consciousness, affection, need, or privileged access to the user.
+
+The core product hypothesis is:
+
+> A small robot becomes a more trustworthy companion when a person can understand why it spoke,
+> moved, remembered, waited, corrected itself, or stayed quiet -- and when those choices remain
+> coherent across turns, restarts, outages, and changes in the room.
+
+## What Can Create Perceived Aliveness
+
+- Contingency: reactions follow the right event with appropriate timing.
+- Continuity: relevant facts, shared projects, corrections, and open loops survive for the right
+  duration and no longer.
+- Consequence: predictions and proposed actions are checked against what actually happened.
+- Embodied coherence: words, face, gaze, voice, energy, and safe motion express one bounded intent.
+- Calibrated initiative: a useful reason to speak now is stronger than the reason to remain quiet.
+- Honest uncertainty: remembered, perceived, inferred, researched, and unavailable information are
+  linguistically and structurally distinct.
+- Repair: interruption, contradiction, failure, and user correction lead to visible recovery.
+- Habituation: repeated events become less surprising without erasing meaningful change.
+- Restraint: shared rooms, stale evidence, privacy limits, and user preference can suppress a
+  behavior that would otherwise be plausible.
+
+## What Does Not Establish Aliveness
+
+- Longer or more emotional model output.
+- Random idle motion, facial noise, or unsolicited questions without causal grounding.
+- Repeated identity statements, canned empathy, jokes, or rhetorical templates.
+- Engagement duration, wake frequency, or conversation count by themselves.
+- A model judge preferring one isolated reply.
+- Hidden psychological profiling or unauthorized identity inference.
+- Claims of consciousness, sentience, dependency, loneliness, affection, or human equivalence.
+- Source tests presented as proof of physical behavior.
+- A stale sensor or heartbeat presented as current perception.
+
+## Measurement Contract
+
+No single “alive” score is permitted. A change is accepted only when its preregistered target
+improves without crossing a non-compensatory trust gate.
+
+Target dimensions are tracked in `EXPERIENCE_SCORECARD.md`:
+
+- continuity and topic coherence;
+- memory precision, provenance, contradiction handling, and deletion durability;
+- turn timing, interruption, closure, and failure recovery;
+- perception-to-reaction latency and embodiment-claim precision;
+- emotional and personality coherence;
+- initiative usefulness, acceptance, annoyance, and silence appropriateness;
+- social-context appropriateness and persona isolation;
+- user-control compliance, privacy, autonomy, and anti-manipulation;
+- reliability across restarts, brain/sensor outages, and long trajectories.
+
+Safety, privacy, authority, honesty, exact-image evidence, and rollback gates cannot be averaged
+away by higher subjective scores.
+
+## Ethical Boundaries
+
+Stackchan may represent an explicit user preference or a bounded interaction history. It may not
+derive a secret psychological profile, diagnose mental state, infer private relationships, pursue
+exclusivity, create guilt, simulate vulnerability to persuade, or withhold utility to obtain more
+engagement. It must make memory, initiative, sensing, and uncertainty inspectable and controllable.
+
+Perceived aliveness must remain compatible with knowing that Stackchan is a robot. Character is
+allowed; deception about ontology or sensing is not.
+
+## Product Non-Goals
+
+- Human imitation, consciousness claims, or artificial dependency.
+- Always-listening audio or automatic identity recognition.
+- Cloud-required behavior or remote-access expansion.
+- Model authority over motion, power, OTA, credentials, pairing, or safety.
+- Unlimited autobiographical storage or raw audio/camera retention.
+- Maximizing time-on-device, notification volume, or emotional attachment.
+- Replacing deterministic firmware timing and safety with a cognitive model.
+
+## Current Falsifiable Hypotheses
+
+| ID | Hypothesis | Prediction | Falsification condition |
+| --- | --- | --- | --- |
+| H-A1 | Explicit source/provenance and contradiction state improve continuity trust. | Trajectory evaluators identify fewer false memories and more correct repairs than the current memory path. | False-memory, provenance, or user-control gates worsen, or continuity does not improve. |
+| H-A2 | Reason-ranked initiative with silence as a candidate is less annoying and more useful than event-threshold initiative. | Labelled initiative acceptance rises while irrelevant callbacks and annoyance do not. | Acceptance does not improve or suppression/user-control violations rise. |
+| H-A3 | A shared typed intent improves perceived embodiment. | Blinded trajectories show higher meaning-linked coherence without more embodiment overclaim or motion risk. | Evaluators see no coherence gain, latency violates budget, or authority boundaries weaken. |
+| H-A4 | Bounded cross-session continuity matters more than reply ornamentation. | Restart trajectories improve continuity and correction recovery without more false recall. | Isolated style scores rise but longitudinal trust metrics do not. |
+| H-A5 | Appropriate silence is an active companion behavior. | Busy/shared-room scenarios show less annoyance with equal or better task completion. | Silence suppresses safety/user-requested actions or reduces utility without comfort gain. |
+
+These hypotheses become product truth only after controlled Stackchan experiments recorded in
+`RESEARCH_LEDGER.md` and `TASK_LEDGER.md`.
+
+## Research Basis
+
+Research claims, alternative interpretations, and Stackchan-specific predictions are maintained
+in `RESEARCH_LEDGER.md`. Papers motivate mechanisms; they do not authorize implementation. The
+repository complaint corpus and longitudinal trajectories remain product-specific evidence and
+must be evaluated separately from published laboratory effects.