feat: pwm-tui ratatui demo proving the real checkpoint locally, no Docker

.github/workflows/release.yml

@@ -143,11 +143,11 @@ jobs:
           lake exe cache get
           lake build
 
-      # Package the publishable crates (pwm-testkit is publish=false). The whole
+      # Package the publishable crates (pwm-testkit and pwm-tui are publish=false). The whole
       # workspace is packaged together so inter-crate deps resolve among the set.
       - name: Package crates
         run: |
-          cargo package --workspace --exclude pwm-testkit --no-verify --locked
+          cargo package --workspace --exclude pwm-testkit --exclude pwm-tui --no-verify --locked
           ls -1 target/package/*.crate
 
       - name: Extract release notes

.github/workflows/tui-smoke.yml

@@ -0,0 +1,76 @@
+# SPDX-License-Identifier: Apache-2.0
+#
+# Headless smoke for the pwm-tui real-checkpoint demo (no Docker). Separate from
+# the merge gate because it downloads external Hugging Face assets and a torch
+# environment. Runs on demand and weekly on main.
+name: TUI Smoke
+
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: "41 5 * * 1"
+
+permissions:
+  contents: read
+
+jobs:
+  tui-headless:
+    name: pwm-tui headless real demo
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    env:
+      NO_COLOR: "1"
+      PWM_CACHE_DIR: ${{ github.workspace }}/.pwm-cache
+    steps:
+      - uses: actions/checkout@v6
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v8
+
+      - name: Build pwm-tui
+        run: cargo build -p pwm-tui --release --locked
+
+      - name: Cold run (downloads pinned assets, exports, proves)
+        run: |
+          set -euo pipefail
+          ./target/release/pwm-tui --headless --json | tee tui-cold.log
+          tail -n 1 tui-cold.log > tui-summary.json
+          python3 - <<'PY'
+          import json
+          s = json.load(open("tui-summary.json"))
+          assert s["ok"] is True, s
+          r = s["report"]
+          assert r["accepted"] is True
+          assert r["weights_root_source"] == "export-bundle"
+          assert float(r["float_error"]) <= float(r["float_tolerance"])
+          assert r["tamper"]["rejected_with"], "tamper must be rejected"
+          print("cold run ok:", r["weights_root"][:16], r["z_out_head"])
+          PY
+
+      - name: Warm run (offline, bundle cache hit, pure Rust)
+        run: |
+          set -euo pipefail
+          ./target/release/pwm-tui --headless --offline --json | tee tui-warm.log
+          grep -q "skipped: bundle cache hit" tui-warm.log
+          tail -n 1 tui-warm.log | python3 -c "import json,sys; s=json.load(sys.stdin); assert s['ok'] is True"
+
+      - name: Cache management commands
+        run: |
+          set -euo pipefail
+          ./target/release/pwm-tui cache path
+          ./target/release/pwm-tui cache ls
+          ./target/release/pwm-tui cache clear --yes
+          ./target/release/pwm-tui --headless --offline > offline-cold.log 2>&1 && exit 1 || true
+          grep -q "offline" offline-cold.log
+
+      - name: Upload logs
+        if: always()
+        uses: actions/upload-artifact@v7
+        with:
+          name: tui-smoke-${{ github.run_id }}
+          path: |
+            tui-cold.log
+            tui-warm.log
+            offline-cold.log
+          if-no-files-found: warn
+          retention-days: 14

AGENTS.md

@@ -41,7 +41,7 @@ non-reproducible attention. There is no proving circuit and no arithmetization.
 
 ## Layout
 
-Five small crates, one trust anchor. The verifier depends on neither the exporter
+Six small crates, one trust anchor. The verifier depends on neither the exporter
 nor any Python or float runtime.
 
 - `pwm-core`: fields (M31 value, Fp61 audit), fixed point, tensors, Merkle
@@ -55,7 +55,12 @@ nor any Python or float runtime.
 - `pwm-verifier`: CPU, `no_std`, float-free. Freivalds-check the linears,
   recompute the rest, check rollout, cost, argmin.
 - `pwm-testkit`: golden vectors, accept and reject suites, the mutation harness,
-  the `pwm` demo CLI (`crates/pwm-testkit/src/bin/pwm.rs`).
+  the `pwm` demo CLI (`crates/pwm-testkit/src/bin/pwm.rs`), and the shared
+  `report` module: the single load -> prove -> verify -> tamper computation that
+  every demo front end renders.
+- `pwm-tui`: the ratatui demo front end. Proves the real checkpoint locally with
+  no Docker, with sha256-pinned checkpoint/dataset/bundle caching and a
+  `--headless --json` mode for CI. Demo only; not published.
 
 ## Run the demo
 
@@ -68,13 +73,22 @@ docker compose --profile compact up --build prover verifier      # tiny two-part
 Without Docker:
 
 ```bash
+cargo run -p pwm-tui --release                                           # REAL checkpoint, TUI, cached
 cargo run -p pwm-testkit --bin pwm --release -- prove-predictor          # synthetic
 cargo run -p pwm-testkit --bin pwm --release -- prove-predictor <bundle> # real weights
 cargo test --workspace                                                   # accept + reject suites
 bash ci/check-local.sh quick                                             # fast local quality loop
 bash ci/check-local.sh full                                              # local merge-gate mirror
 ```
 
+`pwm-tui` mirrors `./demo/run-real.sh` natively: it downloads and pin-verifies
+the checkpoint plus a real episode into a cache (`--cache-dir`, then
+`PWM_CACHE_DIR`, then the platform default), runs the exporter via `uv`
+(network-free, NDJSON progress through `LEWM_JSON_EVENTS`), and proves the real
+weights in-process. A warm bundle cache skips Python entirely. `--offline`,
+`--refresh`, `--re-export`, and `cache path|ls|clear` manage the cache; the
+Docker paths are unchanged.
+
 The `pwm` CLI prints a five-stage pipeline (LOAD, INFER, COMMIT, VERIFY, TAMPER); it
 honors `NO_COLOR` and `--json`.
 

CHANGELOG.md

@@ -13,6 +13,27 @@ entry.
 
 ## [Unreleased]
 
+### Added
+
+- **`pwm-tui`: a ratatui demo that proves the REAL pretrained checkpoint end to
+  end locally, with no Docker** (#200): downloads and sha256-pin-verifies the
+  `quentinll/lewm-pusht` checkpoint and a real `lerobot/pusht` episode into a
+  configurable cache (`--cache-dir` > `PWM_CACHE_DIR` > platform default), runs
+  the PyTorch exporter via `uv` network-free, then proves, verifies, and
+  tamper-checks the real weights in-process with live per-stage progress, the
+  op histogram, the commitments, and the verdicts. A warm bundle cache skips
+  Python entirely; `--offline`, `--refresh`, `--re-export`, `--headless --json`,
+  and `cache path|ls|clear` manage the cache. Exit code 0 means the honest proof
+  was accepted and the forged one rejected.
+- The Python exporter emits NDJSON progress events when `LEWM_JSON_EVENTS` is
+  set (pretty log moves to stderr; the Docker path is unchanged) and accepts
+  `LEWM_PUSHT_PARQUET` / `LEWM_PUSHT_MP4` local file overrides so a caching
+  front end can make the export stage network-free.
+- The `pwm prove-predictor` demo computation moved into the shared
+  `pwm_testkit::report` module (`build_predictor_report`, `PredictorReport`,
+  with phase callbacks); the CLI and the TUI render the same single computation,
+  so their numbers cannot drift.
+
 ## [0.1.0] - 2026-06-11
 
 ### Security