Feat/synthetic only animated map

.dockerignore

@@ -28,19 +28,6 @@ venv/
 .venv/
 .env
 
-# contrail_ml and contrail_flights are excluded entirely — ML training/inference
-# and the OpenSky real-traffic loader are separate concerns with their own heavy
-# deps (scipy/sklearn/xgboost/mlflow; pyopensky/pandas/pyarrow). The server image
-# has one job: run the CP-SAT gRPC service. Both run offline in their own envs.
-contrail_ml/
-contrail_flights/
-
-# contrail_ml artifacts
-mlruns/
-data/processed/
-*.joblib
-*.npz
-
 # Editor / OS noise and local notes
 .DS_Store
 *.ipynb_checkpoints/

.github/workflows/ci.yml

@@ -27,41 +27,11 @@ jobs:
         run: ruff check .
 
       - name: Mypy (type-check)
-        run: mypy contrail_env service contrail_flights
+        run: mypy contrail_env service
 
       - name: Pytest
         run: pytest -q
 
-  ml:
-    # The contrail_ml ISSR model + MLOps lifecycle. Runs the HERMETIC ML tests
-    # (synthetic fallback only — no network, no credentials). Separate job so a
-    # heavy [ml] install never slows the core lint-type-test gate, and the ML
-    # deps never touch the runtime Docker image.
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v4
-
-      - uses: actions/setup-python@v5
-        with:
-          python-version: "3.11"
-
-      - name: Install package with ML + dev extras
-        run: |
-          python -m pip install --upgrade pip
-          pip install -e ".[ml,dev]"
-
-      - name: Generate gRPC stubs
-        run: bash scripts/gen_proto.sh
-
-      - name: Ruff (lint contrail_ml)
-        run: ruff check contrail_ml
-
-      - name: Mypy (type-check contrail_ml)
-        run: mypy contrail_ml
-
-      - name: Pytest (hermetic ML tests)
-        run: pytest tests/test_ml_*.py -q
-
   docker-build:
     runs-on: ubuntu-latest
     steps:
@@ -88,8 +58,8 @@ jobs:
   docker-publish:
     runs-on: ubuntu-latest
     # Only publish on pushes to main — not on pull requests.
-    # The full suite (core + ML) AND the image smoke-test must pass first.
-    needs: [lint-type-test, ml, docker-build]
+    # The full suite AND the image smoke-test must pass first.
+    needs: [lint-type-test, docker-build]
     if: github.event_name == 'push' && github.ref == 'refs/heads/main'
     permissions:
       contents: read

.gitignore

@@ -17,10 +17,5 @@ build/
 # Generated gRPC stubs (regenerated by scripts/gen_proto.sh)
 service/generated/
 
-# contrail_ml — MLflow runs, versioned datasets, trained artifacts (all rebuilt)
-mlruns/
-data/processed/
-*.joblib
-*.npz
-*.dvc
-.dvc/
+# Benchmark output
+results.csv

Dockerfile

@@ -32,12 +32,10 @@ COPY pyproject.toml README.md ./
 COPY contrail_env/ ./contrail_env/
 COPY service/ ./service/
 COPY scripts/ ./scripts/
-# contrail_ml and the quantum solvers (pasqal_analog, xanadu_gbs) are
-# intentionally excluded. This image has one job: run the CP-SAT gRPC server.
-# ML training/inference and quantum sampling are separate concerns that carry
-# heavy deps (scipy/sklearn/xgboost/pulser/strawberryfields) and run in their
-# own environments. Including their source here would ship dead code that can
-# never execute (deps absent) and would crash at request time if called.
+# Only contrail_env + service ship here — the image's one job is the CP-SAT gRPC
+# server. The quantum solver modules (pasqal_analog, xanadu_gbs) ride along in
+# contrail_env and run on their built-in fallbacks; the optional [quantum] SDKs
+# (pulser/strawberryfields) are not installed, keeping the runtime image lean.
 RUN pip install .
 
 # The gRPC stubs are gitignored — generate them from solver.proto at build time.

README.md

@@ -84,7 +84,7 @@ The dashboard has six tabs: live CP-SAT convergence (over ZMQ), the conflict-gra
 topology, QUBO matrix statistics (size, sparsity, penalty constants), the
 chosen-option trade-offs, the quantum benchmark (CP-SAT vs Pasqal vs Xanadu over
 N seeds, with live convergence curves for the BO loop and the GBS sampler), and a
-geographic map — the predicted ISSR risk as a marker overlay on a real Plotly
+geographic map — the ISSR risk as a marker overlay on a real Plotly
 `geo` basemap (country borders / coastlines, drawn with SVG and bundled offline
 vectors, so it needs no WebGL or network) with the chosen vs context routes on top.
 
@@ -94,51 +94,24 @@ The benchmark also runs headless:
 python -m contrail_env.benchmark --flights 4 --seeds 5 --csv results.csv
 ```
 
-## ISSR model — real weather instead of synthetic blobs
+## ISSR field
 
-By default the airspace's contrail zones are synthetic Gaussian blobs. The
-`contrail_ml` package replaces them with a trained model that predicts
-ice-supersaturated regions (ISSRs) from real weather, bias-corrected against
-in-situ IAGOS humidity, and wraps it in a full MLOps lifecycle (versioned data →
-train → calibrate → register with MLflow → serve → monitor).
-
-The model plugs in through the **same `ISSRField` interface** the synthetic
-field uses, so `World` and the QUBO assembly are untouched — you flip a switch:
-
-```python
-from contrail_env import default_european_world
-world = default_european_world(issr_source="ml", issr_kwargs={...})   # vs "synthetic"
-```
-
-Over gRPC the `ScenarioConfig` gained an `issr_source` field (default
-`"synthetic"`, so existing clients are unaffected). Install the extra and try the
-whole serving seam offline:
-
-```
-pip install -e ".[ml]"
-python -m contrail_ml serve-check          # ML ISSR field -> CP-SAT solve
-python -m contrail_ml train --synthetic --no-mlflow   # CV, fit, calibrate, baseline table
-```
-
-The science, the model, and how to read the model-vs-baselines table are in
-[docs/ML.md](docs/ML.md); data sources (IAGOS / ARCO-ERA5 / GFS) in
-[docs/DATA.md](docs/DATA.md). The hermetic tests use a guarded synthetic
-fallback — no network, no credentials.
+The airspace's contrail zones (ice-supersaturated regions) are synthetic
+Gaussian blobs — a controllable obstacle field for the optimizer to route
+around. The field is consumed through a small `ISSRField` interface
+(`rhi_excess`, `is_inside`, `mask_grid`), so the *source* of the field is
+pluggable without touching `World` or the QUBO assembly.
 
 ## Layout
 
 ```
-contrail_env/   synthetic environment, QUBO assembly, CP-SAT solver,
-                quantum pipelines (pasqal_analog, xanadu_gbs, quantum_common,
-                bayes_opt) and the benchmark protocol (benchmark.py)
-contrail_ml/    ISSR model (features, RHiCorrector, calibration), MLflow
-                registry, serving (MLIssrField), monitoring, and the data
-                loaders (IAGOS/ERA5/GFS) — the [ml] extra
+contrail_env/   synthetic environment, ISSR field, geo anchor, QUBO assembly,
+                CP-SAT solver, quantum pipelines (pasqal_analog, xanadu_gbs,
+                quantum_common, bayes_opt) and the benchmark protocol (benchmark.py)
 service/        gRPC service, ZMQ progress streaming, client
 gui/            PyQt6 dashboard
 tests/          environment build, CP-SAT vs brute-force, quantum solvers vs
-                brute-force, benchmark round-trip, gRPC round-trip, and the
-                hermetic contrail_ml suite (test_ml_*.py)
+                brute-force, benchmark round-trip, gRPC round-trip, GUI map panel
 ```
 
 ## Development
@@ -154,11 +127,6 @@ CI runs the same checks on every push and pull request.
 
 ## Roadmap
 
-- **Done:** real ISSR model (`contrail_ml`) replacing the synthetic field, with
-  the train → calibrate → register → serve → monitor MLOps lifecycle (see
-  [docs/ML.md](docs/ML.md)). Next: run it end-to-end on real IAGOS + ERA5 (the
-  loaders are written; they need portal registration) and report the honest
-  real-data comparison table.
 - Run the Pasqal pipeline on real Pulser hardware: needs `[quantum]` extras plus a
   conflict graph that embeds as a valid unit-disk register (auto-detected; the
   built-in simulator is the fallback).

contrail_flights/geo.py → contrail_env/geo.py

@@ -1,20 +1,13 @@
 """
-geo.py — local sim frame <-> real (lon, lat), for the real-flight loader.
+geo.py — Where the synthetic world sits on Earth (local sim frame <-> lon/lat).
 
-This is a DELIBERATE, independent mirror of
-`contrail_ml.features.GeoAnchor`: the brief requires `contrail_flights` to be
-a standalone sibling of `contrail_ml` (no import between the two), so the
-transform is duplicated here rather than shared. The two MUST stay in sync —
-same formula, same default origin — so the predicted ISSR field (placed via
-the contrail_ml anchor) and the real flights (placed via this anchor) agree on
-where things are on the map. `tests/test_geo_transform_roundtrip.py` checks the
-inverse, and a cross-check test guards against drift from the contrail_ml copy.
+The optimizer works entirely in the local 1500x800 km Cartesian box; geography
+only matters for the GUI Map tab, which needs to place that box on a real
+basemap. This module owns that one mapping so the map has a home for it without
+reaching into any heavier package.
 
     lat = origin_lat + y_km / KM_PER_DEG_LAT
     lon = origin_lon + x_km / (KM_PER_DEG_LAT * cos(lat))
-
-Only numpy is needed, so this module imports cleanly in the lean install (the
-[flights] extra is only for the actual OpenSky pull, not the geometry).
 """
 
 from __future__ import annotations
@@ -23,20 +16,26 @@
 
 import numpy as np
 
-# Must match contrail_ml.config.MLConfig defaults so flights and the predicted
-# field share one coordinate system.
-DEFAULT_ORIGIN_LAT = 43.0
-DEFAULT_ORIGIN_LON = -5.0
-KM_PER_DEG_LAT = 111.0
+
+def _maybe_scalar(arr: np.ndarray, like: np.ndarray | float) -> np.ndarray | float:
+    """Return a Python float when the input was scalar, else the array."""
+    if np.isscalar(like) or (isinstance(like, np.ndarray) and like.ndim == 0):
+        return float(arr)
+    return arr
 
 
 @dataclass(frozen=True)
 class GeoAnchor:
-    """Maps the local sim frame (x_km east, y_km north) to/from (lon, lat)."""
+    """Maps the local sim frame (x_km east, y_km north) to geography.
 
-    origin_lat: float = DEFAULT_ORIGIN_LAT
-    origin_lon: float = DEFAULT_ORIGIN_LON
-    km_per_deg_lat: float = KM_PER_DEG_LAT
+    The sim is a flat Cartesian box; the map lives on (lon, lat). One
+    small-angle anchor ties them together so the risk overlay and the routes
+    render over the same place.
+    """
+
+    origin_lat: float = 43.0
+    origin_lon: float = -5.0
+    km_per_deg_lat: float = 111.0
 
     def local_to_geo(
         self, x_km: np.ndarray | float, y_km: np.ndarray | float
@@ -59,8 +58,6 @@ def geo_to_local(
         return _maybe_scalar(x, lon_deg), _maybe_scalar(y, lat_deg)
 
 
-def _maybe_scalar(arr: np.ndarray, like: np.ndarray | float) -> np.ndarray | float:
-    """Return a Python float when the input was scalar, else the array."""
-    if np.isscalar(like) or (isinstance(like, np.ndarray) and like.ndim == 0):
-        return float(arr)
-    return arr
+# The canonical anchor: places the 1500x800 km box over south-west -> central
+# Europe (roughly Madrid to Frankfurt), matching the default world geometry.
+EUROPEAN_ANCHOR = GeoAnchor(origin_lat=43.0, origin_lon=-5.0)

contrail_env/world.py

@@ -178,14 +178,8 @@ def default_european_world(
         - Uniform 3x3 sector grid with capacity 3
 
     issr_source:
-        "synthetic" (default) — the random Gaussian-blob ISSR field, exactly
-            as before, so all existing callers/tests are unaffected.
-        "ml" — a trained model's predicted field (contrail_ml.MLIssrField),
-            which honours the SAME ISSRField interface, so World/qubo need no
-            changes. contrail_ml is imported LAZILY here, so the core install
-            (no [ml] extra) still imports contrail_env fine as long as the ml
-            source is not requested.
-        `issr_kwargs` is forwarded to the chosen ISSR-field factory.
+        "synthetic" (default) — the random Gaussian-blob ISSR field.
+        `issr_kwargs` is forwarded to the ISSR-field factory.
     """
     from .synthetic_issr import random_issr_field
 
@@ -205,14 +199,9 @@ def default_european_world(
             seed=seed,
             **kw,
         )
-    elif issr_source == "ml":
-        # Lazy import: only the "ml" branch pulls in the [ml] extra.
-        from contrail_ml.issr_field import ml_issr_field
-
-        issr = ml_issr_field(**kw)
     else:
         raise ValueError(
-            f"unknown issr_source {issr_source!r} (use 'synthetic' or 'ml')"
+            f"unknown issr_source {issr_source!r} (use 'synthetic')"
         )
 
     from .airspace import uniform_sector_grid