将场加权标定模型接入回放分析

回放按声明的 30 帧加权空载窗口取基线并逐帧保存外推质量，field 模型单独进缓存键，实时入口拒绝。
新增测试覆盖窗口、质量与拒绝路径。全量门禁的 scene3d 两项失败与 studio 访问冲突为既有问题，
已归因并另立 studio-gate-hang。

apps/studio/docs/replay-analysis.md

@@ -38,10 +38,9 @@ zlib level 6。
 `(max − min) / 510`，实测 6.6e-5（该帧峰值 0.034）。这不到深度视图所用 256 级色标的一步，所以
 重建出来的帧**不可能被染成与原帧不同的颜色**。
 
-稀疏存储实测过并否决：每像素 4 B 索引在最坏帧上要 910 B，而整幅量化平面只要 486 B —— 一片几乎
-全零的字节平面在 zlib 下几乎免费，像素索引不是。限制在岛内也实测过并否决：418 B 对 486 B，
-在 4,048 B 的记录里省 5.2 B/帧，代价是重建要依赖有效性掩码 keyframe（box 链路根本没有），
-而且一旦有深度值落在岛外就会**静默丢掉一个测量**。
+稀疏存储实测后否决：最坏帧每像素 4 B 索引要 910 B，整幅量化平面只要 486 B —— 几乎全零的
+字节平面在 zlib 下几乎免费。限制在岛内也否决：418 B 对 486 B，每帧省 5.2 B，却要依赖有效性
+掩码 keyframe（box 链路没有），且岛外深度值会静默丢测量。
 
 `provenance` 随结果存：算法版本、传感器配置及其内容哈希、标定 id、单位、参考帧 id、首个输出
 帧 id、解码偏移、逐通道的"重算/记录"标记（`flow`、`depth` 重算；`forces` 见第 1.1 节）、
@@ -90,7 +89,8 @@ box 链路没有 `valid_mask`，`max_arrows=400` 的上限因此按幅值排序
 `magnitude` 重新推导 `active`，不需要重建结果；只有采样**网格**（`flow_grid_spacing`、
 `flow_edge_margin`、`max_arrows`）变化才会移动候选位置，那份策略记在 provenance 里供比对而
 不塞进键里。SDK 版本同样不在键里（补丁版本不该让键变化），它记在 provenance 的 `versions`，
-由 `AnalysisReader.stale_reasons()` 报出。
+由 `AnalysisReader.stale_reasons()` 报出。场加权实验模型的键分量与逐帧质量见
+[replay-field-experiment.zh.md](replay-field-experiment.zh.md)。
 
 ## 4. 参考帧规则
 

apps/studio/docs/replay-field-experiment.zh.md

@@ -0,0 +1,62 @@
+# Studio 回放：场加权实验模型
+
+用途：reference。范围：Studio 回放分析对 SDK `field_gain/1`（输入 profile
+`hhd_normal_l1_field/1`）的接线、缓存身份、逐帧质量诊断与实时拒绝边界。
+格式、完整性、失效键与播放语义分别由[replay-analysis.md](replay-analysis.md)与
+[replay-analysis-part-b.md](replay-analysis-part-b.md)负责；模型语义见
+[场模型契约](../../../orisys-sdk/docs/field-gain-model.zh.md)，任务范围见
+[实施计划](../../../docs/workflow/field-weighted-calibration-plan.zh.md)。
+
+## 缓存身份
+
+场加权模型是新的分析行为，但不是新的二进制格式。`result_key` 的 `sensor_config` 分量在
+标定文件实际为 `field_gain/1` 时追加 `;field-weighted=1`；spatial/global 模型与旧结果的键
+不变，因此旧回放结果仍能命中。标定内容身份仍来自 schema 3 指纹（含 `pipeline.field`、
+几何与坐标附件），算法版本不因本功能整体升级。
+
+provenance 另存 `field` 块：`profile`、`strategy`、实验来源、`baseline_N`、
+`reference_id`、参与统计帧数、`extrapolated_mean/max` 与**运行时实际**的
+`hhd.downsample`、`hhd.grid_shape`。缓存键只用标定指纹与配置内容哈希，不用声明值假装实际值；
+实际网格只进 provenance。
+
+## 回放 pass
+
+- 打开链后读 `sensor.calibration.input_profile` 判定模型；field 模型只在显式 `cal_path`
+  下使用，profile 默认文件不假定为实验模型。
+- 新模型要求**单 segment 录像**：加权基线无法在段边界重建，多 segment 直接 `ValueError`
+  并留下原因，不静默丢基线。
+- 空载窗口仍需操作员在分析设置里声明（30 帧）；每个 pass 在参考帧 `reset` 之后、该帧
+  `process_chain` 之前调用一次 `declare_unloaded_reference(reference_id="video-unloaded-<id>")`。
+  SDK 累计连续 30 个有效加权 `P` 并取中位数；前 30 帧状态 `zero_unavailable`，
+  没有牛顿值但保留 raw 读数。不得调用 `set_calibration_reference` 传旧 raw 标量零点，
+  也不得把 `centroid_found` 当门槛。
+- 第 30 个有效帧后确认 `baseline_N` 已就绪；视频结束仍未就绪沿用现有 `ValueError` 语义。
+- 逐帧质量取自 `sensor.calibration_result.diagnostics`：`extrapolated_fraction`、
+  `coverage_fraction`、`pending_reference_frames`、`baseline_N`、`reference_id`、
+  `model_fingerprint`。质量是估计证据，不参与 `calibration_status`，不得因此丢弃牛顿值。
+
+## 逐帧存储与读取
+
+场加权质量追加在帧记录最后：`FRAME_FLAG_FIELD` 后接 `extrapolated_fraction` 与
+`coverage_fraction` 两个 float32，沿用 `FRAME_FLAG_FORCE` 的追加先例，格式版本不变。
+`AnalysisReader.frame().field_quality` 返回二元组；旧结果没有该 flag 时返回 `None`，
+上层显示“未知”，不得补零或拒绝结果。`counts.field_quality_frames` 记录带质量帧数。
+
+`ForceSample.from_reading` 继续只凭 raw 有限性判断 `valid`；质量缺失或外推比例高低都不改变
+有效性。`out_of_range` 与负的基线差保留数值。
+
+## 显示与实时拒绝
+
+- `replay_controller` 在曲线来源与提示中使用 provenance 的 `field` 块：实验模型标签、
+  加权基线、平均外推比例；缺字段显示未知。中英 locales 成对新增键。
+- 实时入口未接入该模型：`SdkAcquisitionWorker.start()`/`FingerAcquisitionWorker.start()`
+  与热切换 `_consume_pending_calibration()` 在打开设备前检查显式 `cal_path`，遇
+  schema 3 `field_gain/1` 抛可诊断的 `RuntimeError`；由现有运行错误处理显示。
+  `calibration_service`/实时 UI 若继续重构，保留同一检查函数供调用。
+
+## 验收与限制
+
+回归见 `shared/tests/test_field_calibration_app.py`、`apps/studio/tests/test_field_experiment_display.py`
+及扩展现有四件套：窗口只声明一次、前 30 帧无牛顿值、第 30 帧后就绪、无接触仍有效、
+多 segment 拒绝、质量 round-trip 与旧结果未知、field 键与 spatial 键不同、实时入口拒绝。
+本次不宣称物理精度或在先模型的精度改善；基线窗口是否平稳、外推比例大小都必须随结果如实报告。

apps/studio/src/orisys_studio/replay_controller.py

@@ -42,6 +42,7 @@ from orisys_app_shared.acquisition.image_backend import (
 )
 from orisys_app_shared.i18n import t
 from orisys_app_shared.force.source import AnalysisCurveSource
+from orisys_app_shared.analysis.model import FIELD_PROFILE
 from orisys_app_shared.paths import recordings_dir
 from PySide6.QtCore import QObject, QTimer, Signal
 from PySide6.QtWidgets import QFileDialog, QWidget
@@ -66,6 +67,59 @@ PLAYBACK_TICK_MS = 33
 ReplayHost = ReplayShell
 
 
+def field_experiment_note(session: Any) -> tuple[str, str]:
+    """What a field-weighted result is, as ``(status line, tooltip)``.
+
+    A field result is an **experiment**: its force is an area-weighted integral of the
+    HHD curl-free field with a baseline measured from a declared unloaded window, and
+    part of its support may be extrapolated. The replay has to say that where the
+    operator looks — the status line and the chart's tooltip — using the result's own
+    provenance. A result that carries the field profile but no diagnostics (an older
+    cache, or a pass that could not collect them) says *unknown*: a missing
+    extrapolated fraction is not zero extrapolation.
+
+    Returns ``("", "")`` for anything that is not a field result, so a scalar model's
+    replay is unchanged.
+    """
+    reader = getattr(session, "reader", session)
+    if reader is None:
+        return "", ""
+    provenance = getattr(reader, "provenance", None)
+    field = getattr(provenance, "field", None)
+    if not isinstance(field, dict) or not field:
+        return "", ""
+    if str(field.get("profile") or "") != FIELD_PROFILE:
+        return "", ""
+    frames = int(field.get("frames") or 0)
+    baseline = field.get("baseline_N")
+    extrapolated = field.get("extrapolated_mean")
+    if frames <= 0 or baseline is None or extrapolated is None:
+        return t("chart.field_unknown"), t("chart.field_unknown_tip")
+    hhd = dict(field.get("hhd") or {})
+    grid_shape = hhd.get("grid_shape")
+    grid = (
+        f"{int(grid_shape[0])}x{int(grid_shape[1])}"
+        if isinstance(grid_shape, (list, tuple)) and len(grid_shape) == 2
+        else "?"
+    )
+    downsample = hhd.get("downsample")
+    note = t(
+        "chart.field_experiment",
+        baseline=f"{float(baseline):.3f}",
+        frames=frames,
+        extrapolated=f"{100.0 * float(extrapolated):.1f}",
+    )
+    tip = t(
+        "chart.field_experiment_tip",
+        fingerprint=str(field.get("model_fingerprint") or "?")[:16],
+        strategy=str(field.get("strategy") or "?"),
+        reference=str(field.get("reference_id") or "?"),
+        grid=grid,
+        downsample="?" if downsample is None else str(downsample),
+    )
+    return note, tip
+
+
 @dataclass
 class ReplayActions:
     """The five menu/shortcut entries whose meaning depends on the replay state."""
@@ -1018,6 +1072,18 @@ class ReplayController(QObject):
         # A result just became readable, so it is what the operator is shown — unless a
         # derived view was already on screen (a depth view chosen while the pass ran).
         self.serve_stored_view()
+        # A field-weighted result is an experiment; name its source and quality where
+        # the operator reads the measurement. A scalar result answers ("", "") and
+        # nothing is added.
+        note, tip = field_experiment_note(session)
+        if note:
+            self._host.notify(note)
+            try:
+                self._charts.setToolTip(tip or note)
+            except Exception:  # noqa: BLE001 - a tooltip must not break adoption
+                logger.debug(
+                    "could not attach the field experiment tooltip", exc_info=True
+                )
 
     def serve_stored_view(self) -> bool:
         """Put the cached view on screen when a result becomes readable.

apps/studio/tests/test_field_experiment_display.py

@@ -0,0 +1,72 @@
+"""S4: the replay names the field-weighted experiment a result came from.
+
+The field model is an experiment, and its numbers are estimates: the result records
+its model fingerprint, the declared unloaded reference and the mean extrapolated
+fraction. The replay has to say so — and say "unknown" for a result that predates
+the diagnostics, never 0.
+"""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+from orisys_app_shared.analysis.model import FIELD_PROFILE
+from orisys_app_shared.i18n import set_locale, t
+
+from orisys_studio.replay_controller import field_experiment_note
+
+
+def _session(field: dict) -> SimpleNamespace:
+    return SimpleNamespace(
+        reader=SimpleNamespace(provenance=SimpleNamespace(field=field))
+    )
+
+
+FIELD = {
+    "profile": FIELD_PROFILE,
+    "strategy": "field-weighted-baseline-1",
+    "model_fingerprint": "abcdef0123456789" * 4,
+    "reference_id": "video-unloaded-36",
+    "baseline_N": 0.25,
+    "frames": 35,
+    "extrapolated_mean": 0.3,
+    "extrapolated_max": 0.4,
+    "coverage_mean": 0.95,
+    "hhd": {"downsample": 2, "grid_shape": [150, 140]},
+}
+
+
+def teardown_function():
+    set_locale("zh_CN")
+
+
+def test_a_field_result_names_its_baseline_and_extrapolation():
+    set_locale("zh_CN")
+    note, tip = field_experiment_note(_session(dict(FIELD)))
+    assert note == t(
+        "chart.field_experiment",
+        baseline="0.250",
+        frames=35,
+        extrapolated="30.0",
+    )
+    assert "field-weighted-baseline-1" in tip
+    assert "abcdef0123456789" in tip
+    assert "150x140" in tip or "150 × 140" in tip
+
+    set_locale("en_US")
+    english, _ = field_experiment_note(_session(dict(FIELD)))
+    assert english and english != note
+    assert "30.0" in english
+
+
+def test_a_field_result_without_diagnostics_says_unknown():
+    set_locale("zh_CN")
+    note, tip = field_experiment_note(_session({"profile": FIELD_PROFILE}))
+    assert note == t("chart.field_unknown")
+    assert tip == t("chart.field_unknown_tip")
+
+
+def test_a_scalar_or_missing_result_has_no_field_note():
+    assert field_experiment_note(_session({})) == ("", "")
+    assert field_experiment_note(None) == ("", "")
+    assert field_experiment_note(SimpleNamespace(reader=None)) == ("", "")

apps/studio/tests/test_recording_recalibration.py

@@ -171,3 +171,50 @@ def test_explicit_raw_choice_disables_default_calibration_before_processing(monk
     chain = ChainSpec(sensor_type=family, calibration_disabled=True)
     service_module._OpenChain(chain, video, tmp_path).open()
     assert applied == [None]
+
+
+def test_a_field_model_setup_gets_its_own_key_and_the_unloaded_window(
+    monkeypatch, replay_recording, tmp_path
+):
+    """S4: the field model is a different experiment, so it is a different cache entry.
+
+    The strategy component is added by the analysis service from the chosen file's
+    ``model.input_profile``; the 30-frame declaration travels with the setup. A
+    spatial result's key keeps its old components, so its existing cache still hits.
+    """
+    from orisys_app_shared.analysis import bare_video
+    from orisys_app_shared.analysis import service as service_module
+
+    field_file = tmp_path / "field.json"
+    field_file.write_text(
+        json.dumps(
+            {
+                "schema_version": 3,
+                "model": {
+                    "type": "field_gain/1",
+                    "input_profile": "hhd_normal_l1_field/1",
+                },
+            }
+        ),
+        encoding="utf-8",
+    )
+    monkeypatch.setattr(
+        service_module,
+        "calibration_profile",
+        lambda path: ("field_gain/1", "hhd_normal_l1_field/1"),
+    )
+    monkeypatch.setattr(bare_video, "calibration_identity", lambda path: "field@abc")
+
+    service = AnalysisService(
+        Path(replay_recording),
+        chain=ChainSpec(sensor_type="finger", calibration=field_file),
+        store_force=True,
+        unloaded_reference_frames=30,
+    )
+    facts = service.dependency_report()
+    assert facts["model_profile"] == "hhd_normal_l1_field/1"
+    assert facts["key"].components["sensor_config"].endswith(";field-weighted=1")
+    assert facts["key"].components["calibration"] == "field@abc"
+    assert service.unloaded_reference_frames == 30
+    # The recording itself was not touched by key computation.
+    assert (Path(replay_recording) / "session.json").is_file()

shared/src/orisys_app_shared/acquisition/finger_worker.py

@@ -43,6 +43,7 @@ from orisys_app_shared.acquisition.sdk_worker import (
     _capture_sequence,
     calibration_digest,
     capture_status_of,
+    live_field_model_rejection,
     loaded_calibration_version,
     loaded_path,
     process_chain,
@@ -408,6 +409,9 @@ class FingerAcquisitionWorker(RecordSinkMixin):
     ) -> FixedCameraMode:
         # Non-blocking: a previous session that is stuck opening must never hold
         # the caller (the GUI thread) hostage.
+        rejection = live_field_model_rejection(cal_path)
+        if rejection:
+            raise RuntimeError(rejection)
         self.stop()
         previous = self._thread
         ok, detail = orisys_importable()
@@ -643,6 +647,9 @@ class FingerAcquisitionWorker(RecordSinkMixin):
             session = self._session_id
         if path is None or sensor is None:
             return
+        rejection = live_field_model_rejection(path)
+        if rejection:
+            raise RuntimeError(rejection)
         if not hasattr(sensor, "set_calibration"):
             raise RuntimeError("sensor does not support set_calibration")
         if device_id:

shared/src/orisys_app_shared/acquisition/sdk_worker.py

@@ -95,7 +95,7 @@ def _high_resolution_timer():
         if winmm is not None:
             try:
                 winmm.timeEndPeriod(1)
-            except Exception:  # noqa: BLE001
+            except Exception:  # noqa: BLE001, S110 - a failed timer restore is not fatal
                 pass
 
 # Re-exported so ``from ...sdk_worker import SdkPacket`` keeps working; the
@@ -109,6 +109,7 @@ __all__ = [
     "load_fixed_mode",
     "loaded_calibration_version",
     "loaded_path",
+    "live_field_model_rejection",
     "orisys_importable",
     "process_chain",
     "read_depth",
@@ -394,6 +395,8 @@ def read_forces_from_sensor(sensor, *, config_ref: str | Path | None = None):
         if arr.size == 0:
             depth = None
     calibrated = _sensor_force_calibrated(sensor, config_ref=config_ref)
+    result = getattr(sensor, "calibration_result", None)
+    diagnostics = getattr(result, "diagnostics", None)
     return {
         "normal_force": float(fn),
         "shear_x": float(fx),
@@ -410,6 +413,13 @@ def read_forces_from_sensor(sensor, *, config_ref: str | Path | None = None):
         "invalid_reason": error,
         "raw_available": read_raw,
         "calibration_status": dict(getattr(sensor, "calibration_status", {}) or {}),
+        # Evidence from the calibration model itself (the field model's
+        # extrapolated/coverage fractions, baseline and fingerprint). It is carried,
+        # never interpreted here: whether a value is displayed is the model status's
+        # decision, and quality is not validity.
+        "calibration_diagnostics": dict(diagnostics or {})
+        if isinstance(diagnostics, dict)
+        else {},
     }
 
 
@@ -692,6 +702,41 @@ def loaded_calibration_version(sensor) -> str:
     return calibration_digest(calibration)
 
 
+#: The field-weighted model's input profile. Kept as a literal here because the
+#: acquisition package must not import SDK modules unless a run actually opens one;
+#: ``analysis.model`` owns the app's copy of the same name for the replay path.
+_FIELD_INPUT_PROFILE = "hhd_normal_l1_field/1"
+
+
+def live_field_model_rejection(path: str | Path | None) -> str:
+    """Why a chosen calibration file must not be used for a **live** run, or ``""``.
+
+    The field-weighted model (``field_gain/1`` / ``hhd_normal_l1_field/1``) is a
+    replay-only experiment: it needs a declared 30-frame unloaded window and the
+    same-frame HHD field, neither of which the live acquisition loop supplies. The
+    live entry therefore refuses it by name, before the device is opened, instead of
+    loading it and failing per frame. An unreadable file yields ``""`` — the open
+    path already reports that with its own reason, and this helper must not replace
+    one failure with another.
+    """
+    if path is None or not str(path).strip():
+        return ""
+    try:
+        import orisys.calibration as sdk_calibration
+
+        model = sdk_calibration.load(str(path), strict=True)
+    except Exception:  # noqa: BLE001 - the open path reports an unloadable file itself
+        return ""
+    if str(getattr(model, "input_profile", "") or "") != _FIELD_INPUT_PROFILE:
+        return ""
+    return (
+        "the field-weighted calibration model (hhd_normal_l1_field/1) is replay-only: "
+        "it needs a declared 30-frame unloaded window and the same-frame HHD field, "
+        "which the live chain does not provide. Choose a scalar calibration for live "
+        f"operation ({Path(str(path)).name})."
+    )
+
+
 
 class SdkAcquisitionWorker(RecordSinkMixin):
     """Background loop: Sensor.get_img + deformation + force readout."""
@@ -838,6 +883,9 @@ class SdkAcquisitionWorker(RecordSinkMixin):
         # Non-blocking: a previous session that is stuck opening the device must
         # never hold the caller (the GUI thread) hostage.
         del device_id  # box/mini open path has no Finger-style serial check
+        rejection = live_field_model_rejection(cal_path)
+        if rejection:
+            raise RuntimeError(rejection)
         self.stop()
         previous = self._thread
         ok, detail = orisys_importable()
@@ -1023,6 +1071,9 @@ class SdkAcquisitionWorker(RecordSinkMixin):
             session = self._session_id
         if path is None or sensor is None:
             return
+        rejection = live_field_model_rejection(path)
+        if rejection:
+            raise RuntimeError(rejection)
         if not hasattr(sensor, "set_calibration"):
             raise RuntimeError("sensor does not support set_calibration")
         sensor.set_calibration(path)
@@ -1050,7 +1101,7 @@ class SdkAcquisitionWorker(RecordSinkMixin):
             return
         try:
             sensor.disconnect()
-        except Exception:
+        except Exception:  # noqa: BLE001, S110 - teardown must not mask the pass result
             pass
 
     def _close_all_sensors(self) -> None:

shared/src/orisys_app_shared/analysis/model.py

@@ -57,6 +57,17 @@ differently now. This is the one channel whose ``Provenance.sources`` entry is
 ``recomputed`` on the *synthetic* path and ``recorded`` beside a real ``forces.csv``:
 the two must never be confused, which is what the field is for.
 
+**The field model's quality is stored beside the force, not inside it.** The
+field-weighted model is an experiment: its estimate is an area-weighted integral of the
+HHD curl-free field over a mapped gain, and part of that support may be extrapolated
+from the convex hull boundary. The SDK reports that as an extrapolated fraction; the
+pass stores it per frame behind :data:`FRAME_FLAG_FIELD` with an 8-byte
+:data:`FIELD_PAYLOAD` (extrapolated, coverage). It is *evidence about the estimate*,
+never a validity statement: a frame with a non-zero extrapolated fraction still has a
+newton value. A frame without the flag has **unknown** quality (a scalar model, an
+older result) and must be displayed that way, not as 0. The total per-frame cost is
+8 B against the 4,048 B a frame already costs.
+
 Sparse storage was measured and rejected: 4-byte indices per non-zero pixel cost
 910 B for the worst non-empty frame against 486 B for the whole quantised map, because
 a mostly-zero byte plane is nearly free for zlib while pixel indices are not. Island
@@ -92,9 +103,14 @@ __all__ = [
     "CODEC_ZLIB",
     "DEPTH_BITS",
     "DEPTH_HEADER",
+    "FIELD_KEY_SUFFIX",
+    "FIELD_PAYLOAD",
+    "FIELD_PROFILE",
+    "FIELD_STRATEGY",
     "FORCE_PAYLOAD",
     "FRAME_FLAG_DEPTH",
     "FRAME_FLAG_DROPPED",
+    "FRAME_FLAG_FIELD",
     "FRAME_FLAG_FORCE",
     "FRAME_FLAG_VALIDITY",
     "FRAME_HEADER",
@@ -145,6 +161,20 @@ ALGORITHM_VERSION = "flow-hhd-contact-2"
 #: unit :class:`~orisys_app_shared.scene3d.visual_policy.FlowSamples` documents.
 UNIT = "uv_px"
 
+#: The field-weighted model's input profile (SDK calibration contract). The app only
+#: names it: the model is loaded and run by the SDK, and this string is what tells the
+#: pass that the unloaded reference is a 30-frame weighted window rather than a scalar
+#: raw zero. It also keys the result, because that window is a property of this video.
+FIELD_PROFILE = "hhd_normal_l1_field/1"
+#: Cache-key component appended to ``sensor_config`` when the chosen calibration is the
+#: field-weighted model. A spatial result and a field result must never share a key:
+#: the force on the same footage is a different experiment with a different baseline.
+#: It is added only for an explicitly chosen file whose ``model.input_profile`` is
+#: :data:`FIELD_PROFILE`; the scalar models' keys are unchanged.
+FIELD_KEY_SUFFIX = ";field-weighted=1"
+#: Version of the app's field strategy, written to the provenance for display/audit.
+FIELD_STRATEGY = "field-weighted-baseline-1"
+
 #: ``frame_id``, ``t_rel_s``, grid keyframe id, flags, candidate count, depth
 #: non-zero count, depth minimum, depth maximum. The three trailing pad bytes keep
 #: the value arrays 8-byte aligned inside a record.
@@ -160,6 +190,12 @@ FRAME_FLAG_DEPTH = 0x04
 #: pass was not asked to collect it (a recording, whose forces live in its own
 #: ``forces.csv``) or the chain exposed none.
 FRAME_FLAG_FORCE = 0x08
+#: The frame carries the field model's quality payload (:data:`FIELD_PAYLOAD`),
+#: appended after the force payload. This is *evidence about the estimate* — how much
+#: of its support was extrapolated — not a validity statement: the newton value is
+#: still governed by the calibration status and must not be dropped because the
+#: extrapolated fraction is non-zero.
+FRAME_FLAG_FIELD = 0x10
 
 #: The force payload: FN, FX, FY in newtons, then the raw triplet, then one byte of
 #: validity and one of "a calibration curve was applied". Appended **last** in the
@@ -168,6 +204,13 @@ FRAME_FLAG_FORCE = 0x08
 #: always ignored trailing bytes.
 FORCE_PAYLOAD = struct.Struct("<ffffffBB2x")
 
+#: Field quality: extrapolated fraction, coverage fraction (both 0..1). Appended after
+#: the force payload like the force payload is appended after depth: a reader that
+#: does not know :data:`FRAME_FLAG_FIELD` still decodes every earlier field. A frame
+#: without the flag has *unknown* quality (an older result, or a scalar model) — not
+#: zero extrapolation.
+FIELD_PAYLOAD = struct.Struct("<ff")
+
 #: magic, kind, codec, count, height, width, flags, reserved, raw bytes, digest
 KEYFRAME_HEADER = struct.Struct("<4sBBIIIBBQ8s")
 KEYFRAME_MAGIC = b"ORAK"
@@ -515,6 +558,11 @@ class ForceSample:
     #: The calibration model's per-axis status for this frame, as ``(axis, status)``
     #: pairs. Kept in memory for the pass's counters; not part of the stored payload.
     status: tuple[tuple[str, str], ...] = field(default=(), compare=False)
+    #: The calibration model's own evidence for this frame (the field model's
+    #: extrapolated/coverage fractions, baseline and fingerprint). Kept in memory for
+    #: the pass to store as quality; not part of the force payload, and never a reason
+    #: to drop a finite newton value.
+    diagnostics: dict[str, Any] = field(default_factory=dict, compare=False)
 
     @property
     def any_finite(self) -> bool:
@@ -537,6 +585,23 @@ class ForceSample:
     def raw(self) -> tuple[float, float, float]:
         return (self.normal_raw, self.shear_x_raw, self.shear_y_raw)
 
+    def field_quality(self) -> tuple[float, float] | None:
+        """This frame's ``(extrapolated_fraction, coverage_fraction)``, or ``None``.
+
+        Only the field model reports these. ``None`` means *unknown* (an older result,
+        a scalar model, or a frame whose model gave no field evidence) and must never
+        be displayed as 0.
+        """
+        extrapolated = self.diagnostics.get("extrapolated_fraction")
+        coverage = self.diagnostics.get("coverage_fraction")
+        try:
+            values = (float(extrapolated), float(coverage))
+        except (TypeError, ValueError):
+            return None
+        if not all(value == value and abs(value) != float("inf") for value in values):
+            return None
+        return tuple(min(1.0, max(0.0, value)) for value in values)  # type: ignore[return-value]
+
     def values(self, unit: str = "raw") -> tuple[float, float, float]:
         """The triplet in ``unit``: newtons when a calibration was applied, else raw.
 
@@ -565,6 +630,7 @@ class ForceSample:
                 return _NAN
             return float(reading.get(key, _NAN))
 
+        diagnostics = reading.get("calibration_diagnostics")
         return cls(
             normal=newton("fn", "normal_force"),
             shear_x=newton("fx", "shear_x"),
@@ -575,6 +641,7 @@ class ForceSample:
             valid=bool(reading.get("valid", True)),
             calibrated=bool(reading.get("force_calibrated", False)),
             status=tuple(sorted(statuses.items())),
+            diagnostics=dict(diagnostics or {}),
         )
 
     def encode(self) -> bytes:
@@ -628,8 +695,9 @@ def encode_frame(
     depth_nonzero: int = 0,
     dropped: bool = False,
     force: ForceSample | None = None,
+    field_quality: Sequence[float] | None = None,
 ) -> bytes:
-    """One frame record: a fixed header, then the value arrays, depth and force.
+    """One frame record: a fixed header, then the value arrays, depth, force and quality.
 
     Arrays are little-endian regardless of host order, so a result written on one
     machine reads back identically on another. ``dx``/``dy``/``magnitude`` are stored
@@ -645,6 +713,12 @@ def encode_frame(
     ``force`` is the chain's own sample for this frame, or ``None`` when the pass did not
     collect one (``FRAME_FLAG_FORCE``). It is written last, after every field an older
     reader knows, so the record stays readable by a build without this flag.
+
+    ``field_quality`` is the field model's ``(extrapolated_fraction, coverage_fraction)``
+    for this frame, or ``None`` when the model reported none (a scalar model, an older
+    result, or a frame without field evidence). It is appended *after* the force payload
+    for the same reason the force payload follows depth: an older reader never parses
+    past a field it knows, and a chunk's offset table delimits the record.
     """
     dx = np.ascontiguousarray(dx, dtype="<f4").reshape(-1)
     dy = np.ascontiguousarray(dy, dtype="<f4").reshape(-1)
@@ -660,6 +734,16 @@ def encode_frame(
         raise ValueError(f"grid id {grid_id} does not fit the frame header")
     if count > 0xFFFF:
         raise ValueError(f"{count} candidates do not fit the frame header")
+    quality_payload = b""
+    if field_quality is not None:
+        values = tuple(float(value) for value in field_quality)
+        if len(values) != 2 or not all(value == value for value in values):
+            raise ValueError(
+                "field quality must be two finite fractions (extrapolated, coverage)"
+            )
+        quality_payload = FIELD_PAYLOAD.pack(
+            min(1.0, max(0.0, values[0])), min(1.0, max(0.0, values[1]))
+        )
     flags = 0
     if validity_digest is not None:
         flags |= FRAME_FLAG_VALIDITY
@@ -667,6 +751,8 @@ def encode_frame(
         flags |= FRAME_FLAG_DROPPED
     if force is not None:
         flags |= FRAME_FLAG_FORCE
+    if quality_payload:
+        flags |= FRAME_FLAG_FIELD
     depth_payload = b""
     depth_low = depth_high = 0.0
     if depth is not None:
@@ -696,6 +782,7 @@ def encode_frame(
     parts.append(depth_payload)
     if force is not None:
         parts.append(force.encode())
+    parts.append(quality_payload)
     return b"".join(parts)
 
 
@@ -720,6 +807,9 @@ class DecodedFrame:
     #: ``None`` when the frame carried no force (an older result, a recording whose forces
     #: live in its own ``forces.csv``, or a chain that produced none).
     force: ForceSample | None = None
+    #: ``None`` when the frame carried no field quality (an older result, a scalar model,
+    #: or no field evidence). Not the same fact as "the extrapolated fraction was 0".
+    field_quality: tuple[float, float] | None = None
 
     def depth_map(self) -> np.ndarray | None:
         """This frame's depth map, or ``None`` when it had no depth channel."""
@@ -771,6 +861,14 @@ def decode_frame(
         if len(view) < cursor + FORCE_PAYLOAD.size:
             raise ValueError("a frame says it carries force but is shorter than one sample")
         force = ForceSample.decode(view, cursor)
+        cursor += FORCE_PAYLOAD.size
+    field_quality = None
+    if flags & FRAME_FLAG_FIELD:
+        if len(view) < cursor + FIELD_PAYLOAD.size:
+            raise ValueError("a frame says it carries field quality but is shorter than it")
+        field_quality = tuple(
+            float(value) for value in FIELD_PAYLOAD.unpack_from(view, cursor)
+        )
     return DecodedFrame(
         frame_id=int(frame_id),
         t_rel_s=float(t_rel),
@@ -785,6 +883,7 @@ def decode_frame(
         validity_digest=validity,
         depth=depth,
         force=force,
+        field_quality=field_quality,
     )
 
 
@@ -795,18 +894,22 @@ def frame_payload_bytes(
     with_depth: bool = False,
     depth_stored: int = 0,
     with_force: bool = False,
+    with_field_quality: bool = False,
 ) -> int:
     """Stored bytes of one frame record at ``candidates`` sample locations.
 
     ``with_depth`` adds the 6-byte depth header every frame that carries a channel pays;
     ``depth_stored`` adds a payload of that many bytes on top, for a caller sizing the
-    memory bound rather than the common case. ``with_force`` adds the fixed force sample.
+    memory bound rather than the common case. ``with_force`` adds the fixed force sample;
+    ``with_field_quality`` adds the fixed field quality payload.
     """
     total = FRAME_HEADER.size + int(candidates) * 12 + (int(candidates) + 7) // 8
     if with_validity:
         total += 8
     if with_force:
         total += FORCE_PAYLOAD.size
+    if with_field_quality:
+        total += FIELD_PAYLOAD.size
     if with_depth:
         total += DEPTH_HEADER.size + max(0, int(depth_stored))
     return total
@@ -872,6 +975,11 @@ class FramePayload:
     #: which is why the provenance — not this field — is what a chart must label a trace
     #: from.
     force: ForceSample | None = None
+    #: The field model's ``(extrapolated_fraction, coverage_fraction)`` for this frame,
+    #: or ``None`` when the result carries none (a scalar model, an older result). This
+    #: is quality evidence, not validity: a finite newton value with a non-zero
+    #: extrapolated fraction is still a value.
+    field_quality: tuple[float, float] | None = None
 
     @property
     def has_depth(self) -> bool:
@@ -928,6 +1036,7 @@ class FramePayload:
             depth_nonzero=self.depth_nonzero,
             dropped=self.dropped,
             force=self.force,
+            field_quality=self.field_quality,
         )
 
 
@@ -1035,9 +1144,20 @@ class Provenance:
     #: that time is not throughput.
     open_seconds: float = 0.0
     warnings: list[str] = field(default_factory=list)
+    #: True when the pass reached the end of every declared segment and committed.
     complete: bool = False
     stopped_reason: str = ""
     counts: dict[str, Any] = field(default_factory=dict)
+    #: The field-weighted experiment behind this result, when it has one. Keys:
+    #: ``profile``/``strategy`` (what ran), ``model_fingerprint`` (content identity of
+    #: the model), ``reference_id``/``baseline_N`` (the declared unloaded window), and
+    #: ``frames``/``extrapolated_mean``/``extrapolated_max``/``coverage_mean`` (per-frame
+    #: quality statistics) plus ``hhd`` (effective grid shape and downsample the
+    #: weighted sum ran on). Empty for a scalar model or an older result; a consumer
+    #: must display that as *unknown*, never as zero extrapolation.
+    #: Declared last on purpose: an attribute named ``field`` would shadow the
+    #: ``dataclasses.field`` helper for every attribute after it in this class body.
+    field: dict[str, Any] = field(default_factory=dict)
 
     def to_dict(self) -> dict[str, Any]:
         return dict(self.__dict__)

shared/src/orisys_app_shared/analysis/reader.py

@@ -533,12 +533,22 @@ class AnalysisReader:
             depth_max=depth_max,
             depth_stored_bytes=depth_stored,
             force=decoded.force,
+            field_quality=decoded.field_quality,
         )
 
     def force(self, frame_id: int) -> ForceSample | None:
         """One frame's stored force sample, or ``None`` when the frame carries none."""
         return self.frame(int(frame_id)).force
 
+    def field_quality(self, frame_id: int) -> tuple[float, float] | None:
+        """One frame's ``(extrapolated_fraction, coverage_fraction)``, or ``None``.
+
+        ``None`` is the truthful answer for a result that predates the field-model
+        diagnostics (or for a scalar model): the extrapolated fraction was not
+        recorded, which is not the same as zero. Display it as unknown.
+        """
+        return self.frame(int(frame_id)).field_quality
+
     @property
     def has_force(self) -> bool:
         """Whether *any* frame in this result carries a force sample.

shared/src/orisys_app_shared/analysis/service.py

@@ -61,6 +61,9 @@ from orisys_app_shared.analysis.cache import (
 )
 from orisys_app_shared.analysis.model import (
     ALGORITHM_VERSION,
+    FIELD_KEY_SUFFIX,
+    FIELD_PROFILE,
+    FIELD_STRATEGY,
     KEYFRAME_MASK,
     ForceSample,
     Provenance,
@@ -86,6 +89,7 @@ __all__ = [
     "ChainSpec",
     "LiveFrame",
     "Progress",
+    "calibration_profile",
     "chain_spec_for",
     "recording_config_sha256",
     "resolve_finger_export_dir",
@@ -136,6 +140,116 @@ def recording_config_sha256(config_ref: str | Path | None) -> str:
     return ""
 
 
+def calibration_profile(path: str | Path | None) -> tuple[str, str]:
+    """``(model.type, model.input_profile)`` of a calibration file, or ``("", "")``.
+
+    A plain JSON lookup on purpose: the result key has to be computed *before* anything
+    opens the SDK (a cache hit is a file read, and a pass that turns out to be cached
+    must not pay for a chain open). Schema 1/2 documents have no ``model`` block and
+    answer ``("", "")``, which is the same answer an unreadable file gives — the open
+    path reports that failure with its own reason.
+    """
+    if path is None:
+        return ("", "")
+    try:
+        data = json.loads(Path(str(path)).read_text(encoding="utf-8-sig"))
+    except (OSError, ValueError):
+        return ("", "")
+    if not isinstance(data, dict):
+        return ("", "")
+    model = data.get("model")
+    if not isinstance(model, dict):
+        return ("", "")
+    return (
+        str(model.get("type") or ""),
+        str(model.get("input_profile") or ""),
+    )
+
+
+def _chain_raw_sensor(chain: Any) -> Any:
+    """The SDK ``Sensor`` behind a chain (a finger chain wraps it in a tactile one)."""
+    return getattr(chain.sensor, "raw_sensor", chain.sensor)
+
+
+def _chain_field_model(chain: Any) -> Any:
+    """The loaded calibration model when it is the field-weighted one, else ``None``."""
+    model = getattr(_chain_raw_sensor(chain), "calibration", None)
+    if str(getattr(model, "input_profile", "") or "") != FIELD_PROFILE:
+        return None
+    return model
+
+
+def _field_diagnostics(sensor: Any) -> dict[str, Any]:
+    """The calibration model's diagnostics for the frame it just answered.
+
+    The SDK exposes them as ``Sensor.calibration_result.diagnostics``; a sensor that
+    has no model (or a model that reports none) answers ``{}``. Missing diagnostics
+    mean *unknown quality*, never zero extrapolation.
+    """
+    result = getattr(sensor, "calibration_result", None)
+    diagnostics = getattr(result, "diagnostics", None)
+    return dict(diagnostics) if isinstance(diagnostics, dict) else {}
+
+
+def _hhd_identity(sensor: Any) -> dict[str, Any]:
+    """The HHD grid the weighted sum actually ran on, for the result's provenance.
+
+    Read from the open sensor (the effective values), not from the configuration
+    document: the SDK may accept a calibration whose declared ``hhd_downsample``
+    differs from the runtime one, and the area factor ``A`` depends on the grid that
+    actually ran.
+    """
+    identity: dict[str, Any] = {}
+    downsample = getattr(sensor, "_hhd_downsample", None)
+    if downsample is not None:
+        try:
+            identity["downsample"] = int(downsample)
+        except (TypeError, ValueError):
+            pass
+    for key, attribute in (("grid_shape", "_hhd_grid_shape"), ("reference_shape", "_hhd_ref_shape")):
+        value = getattr(sensor, attribute, None)
+        if value is None:
+            continue
+        try:
+            identity[key] = [int(size) for size in value]
+        except (TypeError, ValueError):
+            continue
+    return identity
+
+
+def _accept_field_diagnostics(
+    state: _PassState, sensor: Any, *, zero_frames: int
+) -> None:
+    """Keep this frame's weighted-window status and quality statistics in the state."""
+    diagnostics = _field_diagnostics(sensor)
+    if zero_frames and state.field_reference_id and state.field_baseline is None:
+        baseline = diagnostics.get("baseline_N")
+        if baseline is not None and np.isfinite(float(baseline)):
+            state.field_baseline = float(baseline)
+        pending = diagnostics.get("pending_reference_frames")
+        if pending is not None:
+            state.provenance.counts["field_pending_reference_frames"] = int(pending)
+    extrapolated = diagnostics.get("extrapolated_fraction")
+    if extrapolated is not None:
+        try:
+            value = min(1.0, max(0.0, float(extrapolated)))
+        except (TypeError, ValueError):
+            value = None
+        if value is not None:
+            state.field_extrapolated_sum += value
+            state.field_extrapolated_max = max(state.field_extrapolated_max, value)
+            state.field_extrapolated_count += 1
+    coverage = diagnostics.get("coverage_fraction")
+    if coverage is not None:
+        try:
+            value = min(1.0, max(0.0, float(coverage)))
+        except (TypeError, ValueError):
+            value = None
+        if value is not None:
+            state.field_coverage_sum += value
+            state.field_coverage_count += 1
+
+
 def resolve_finger_export_dir(config_ref: str | Path) -> Path | None:
     """``assets/30fps-0703`` beside a finger ``sensor.json``, when it is complete.
 
@@ -509,6 +623,18 @@ class _PassState:
     frames_expected: int = 0
     frames_written: int = 0
     started: float = 0.0
+    #: The field-weighted experiment this pass ran, when the loaded model is
+    #: ``field_gain/1``. Empty/false for every scalar model.
+    field_profile: str = ""
+    field_fingerprint: str = ""
+    field_reference_id: str = ""
+    field_baseline: float | None = None
+    field_hhd: dict[str, Any] = field(default_factory=dict)
+    field_extrapolated_sum: float = 0.0
+    field_extrapolated_max: float = 0.0
+    field_extrapolated_count: int = 0
+    field_coverage_sum: float = 0.0
+    field_coverage_count: int = 0
 
 
 class AnalysisService:
@@ -631,6 +757,14 @@ class AnalysisService:
             from orisys_app_shared.analysis.bare_video import calibration_identity
 
             calibration = calibration_identity(self.chain.calibration)
+        # The chosen file's model, read as JSON (never by opening the SDK): the field
+        # model's unloaded baseline is measured from *this* video's frames, so a field
+        # result and a scalar result on the same footage must not share a cache entry.
+        model_type, model_profile = (
+            ("", "")
+            if self.chain.calibration_disabled
+            else calibration_profile(self.chain.calibration)
+        )
         zero_frames = self._zero_reference_frames()
         key = result_key(
             recording=identity,
@@ -638,6 +772,7 @@ class AnalysisService:
                 self.chain.identity()
                 + (";unloaded-reference=30" if zero_frames else "")
                 + (";force=recomputed" if self.store_force else "")
+                + (FIELD_KEY_SUFFIX if model_profile == FIELD_PROFILE else "")
             ),
             calibration=calibration,
             reference_frame=self.reference_frame,
@@ -650,6 +785,8 @@ class AnalysisService:
             "segment_sizes": sizes,
             "recording_identity": identity,
             "calibration_id": calibration,
+            "model_type": model_type,
+            "model_profile": model_profile,
             "key": key,
         }
 
@@ -680,6 +817,63 @@ class AnalysisService:
             f"frames ({span}); a reset that cannot be applied is not a result"
         )
 
+    def _require_field_single_segment(self, facts: dict[str, Any]) -> None:
+        """Refuse a field-model pass over more than one segment.
+
+        The field model's baseline is one declared, consecutive 30-frame window of the
+        *same* stream. A segment boundary reopens the chain (and the tracker), so a
+        window that crossed it would be a different measurement wearing the same
+        reference id. Refusing is the only truthful answer; silently re-declaring would
+        discard the first segment's baseline and silently rebuilding it would invent
+        one.
+        """
+        if facts.get("model_profile") != FIELD_PROFILE:
+            return
+        segments = list(facts.get("segments") or ())
+        if len(segments) <= 1:
+            return
+        raise ValueError(
+            "field_gain/1 requires a single-segment recording: its weighted unloaded "
+            "baseline is one declared window of the same stream and cannot be rebuilt "
+            f"across a segment boundary ({len(segments)} segments in {self.recording})"
+        )
+
+    @staticmethod
+    def _field_provenance(state: _PassState) -> dict[str, Any]:
+        """The field-weighted experiment behind this result, as stored evidence.
+
+        Kept even for an abandoned pass: a cancelled analysis still has to say which
+        model it was running and whether its unloaded window completed.
+        """
+        extrapolated_mean = (
+            state.field_extrapolated_sum / state.field_extrapolated_count
+            if state.field_extrapolated_count
+            else None
+        )
+        coverage_mean = (
+            state.field_coverage_sum / state.field_coverage_count
+            if state.field_coverage_count
+            else None
+        )
+        return {
+            "profile": state.field_profile,
+            "strategy": FIELD_STRATEGY,
+            "model_fingerprint": state.field_fingerprint,
+            "reference_id": state.field_reference_id,
+            "baseline_N": state.field_baseline,
+            "frames": state.field_extrapolated_count,
+            "extrapolated_mean": (
+                None if extrapolated_mean is None else round(extrapolated_mean, 6)
+            ),
+            "extrapolated_max": (
+                round(state.field_extrapolated_max, 6)
+                if state.field_extrapolated_count
+                else None
+            ),
+            "coverage_mean": None if coverage_mean is None else round(coverage_mean, 6),
+            "hhd": dict(state.field_hhd),
+        }
+
     # ------------------------------------------------------------------------ pass
     def recompute(self) -> Iterator[LiveFrame]:
         """Re-run the chain over every segment and yield what it computes now.
@@ -692,6 +886,7 @@ class AnalysisService:
         key: ResultKey = facts["key"]
         grouped = self.rows_by_segment(facts["info"].frames)
         self._require_reference(grouped)
+        self._require_field_single_segment(facts)
         working = self._make_working_directory(key)
         state = _PassState(
             provenance=self._provenance(key, facts),
@@ -724,6 +919,7 @@ class AnalysisService:
         # Checked before anything is written, so a caller error does not leave a
         # directory behind at all.
         self._require_reference(grouped)
+        self._require_field_single_segment(facts)
         working = self._make_working_directory(key)
         writer = AnalysisWriter(
             self.cache.directory(key), key=key, chunk_frames=self.chunk_frames
@@ -762,6 +958,10 @@ class AnalysisService:
             state.reasons.append(f"{type(exc).__name__}: {exc}")
         finally:
             self._discard_working_directory(working)
+        if state.field_profile:
+            # Written for complete *and* abandoned results: a cancelled field pass still
+            # has to name its model and whether the unloaded window had completed.
+            provenance.field = self._field_provenance(state)
         if (
             not state.reasons
             and self.reference_frame is not None
@@ -784,6 +984,9 @@ class AnalysisService:
                 writer.bytes_written / state.frames_written, 2
             )
         counts["open_seconds"] = provenance.open_seconds
+        if state.field_profile:
+            counts["field_frames"] = state.field_extrapolated_count
+            counts["field_baseline_ready"] = state.field_baseline is not None
         provenance.decode_order_offsets = dict(state.offsets)
         distinct = sorted(set(state.offsets.values()))
         provenance.decode_order_offset = distinct[0] if len(distinct) == 1 else None
@@ -843,6 +1046,25 @@ class AnalysisService:
                 provenance.warnings.append(f"segment_frame_gap:{number}")
             with _OpenChain(self.chain, path, state.working) as chain:
                 zero_samples = []
+                raw_sensor = _chain_raw_sensor(chain)
+                field_model = _chain_field_model(chain)
+                field_weighted = field_model is not None
+                if field_weighted:
+                    # The baseline is one window of the same stream; a second segment
+                    # would reopen the tracker and could only re-declare it. Refuse
+                    # rather than silently keep the first window or rebuild a new one.
+                    if len(state.segments) > 1:
+                        raise ValueError(
+                            "field_gain/1 requires a single-segment recording: its "
+                            "weighted unloaded baseline is one declared window of the "
+                            "same stream and cannot be rebuilt across a segment boundary"
+                        )
+                    if not state.field_profile:
+                        state.field_profile = FIELD_PROFILE
+                        state.field_fingerprint = str(
+                            getattr(field_model, "fingerprint", "") or ""
+                        )
+                        state.field_hhd = _hhd_identity(raw_sensor)
                 offset, measured = self._segment_offset(chain, number, state.offsets)
                 state.offsets[str(number)] = offset
                 state.sources.add("measured" if measured else "declared")
@@ -888,25 +1110,79 @@ class AnalysisService:
                         provenance.counts["reset_count"] = (
                             provenance.counts.get("reset_count", 0) + 1
                         )
+                    if (
+                        field_weighted
+                        and zero_frames
+                        and not state.field_reference_id
+                        and (
+                            self.reference_frame is None
+                            or int(row.frame_id) >= int(self.reference_frame)
+                        )
+                    ):
+                        # The tracker reset cleared the model's session too, so the
+                        # declaration must follow it and *precede* this frame's chain
+                        # call: otherwise the first frame of the window is consumed
+                        # before the model knows a window started.
+                        declare = getattr(raw_sensor, "declare_unloaded_reference", None)
+                        if not callable(declare):
+                            raise ValueError(
+                                "this SDK cannot declare the field model's weighted "
+                                "unloaded window"
+                            )
+                        state.field_reference_id = (
+                            f"video-unloaded-{int(row.frame_id)}"
+                        )
+                        declare(reference_id=state.field_reference_id)
+                        provenance.counts["field_reference_frames"] = zero_frames
                     depth, error = process_chain(chain.sensor)
                     if error:
                         provenance.warnings.append(f"chain:{error}")
                         state.reasons.append(f"chain_failed:{row.frame_id}")
                         return
-                    if zero_frames and (self.reference_frame is None or row.frame_id >= self.reference_frame) and len(zero_samples) < zero_frames:
-                        sensor = getattr(chain.sensor, "raw_sensor", chain.sensor)
-                        model = getattr(sensor, "calibration", None)
-                        if getattr(model, "schema_version", None) != 3:
-                            raise ValueError("unloaded reference requires a schema 3 calibration")
-                        keys = (sensor.info.FNORMAL_RAW, sensor.info.FSHEARX_RAW, sensor.info.FSHEARY_RAW)
-                        values = sensor.read_info(*keys)
-                        if getattr(sensor, "centroid_found", False) or not np.all(np.isfinite(values)):
-                            raise ValueError("unloaded reference contains contact or invalid readings")
+                    if field_weighted:
+                        _accept_field_diagnostics(
+                            state, raw_sensor, zero_frames=zero_frames
+                        )
+                    if (
+                        not field_weighted
+                        and zero_frames
+                        and (
+                            self.reference_frame is None
+                            or row.frame_id >= self.reference_frame
+                        )
+                        and len(zero_samples) < zero_frames
+                    ):
+                        if getattr(raw_sensor, "calibration", None) is None or getattr(
+                            raw_sensor.calibration, "schema_version", None
+                        ) != 3:
+                            raise ValueError(
+                                "unloaded reference requires a schema 3 calibration"
+                            )
+                        keys = (
+                            raw_sensor.info.FNORMAL_RAW,
+                            raw_sensor.info.FSHEARX_RAW,
+                            raw_sensor.info.FSHEARY_RAW,
+                        )
+                        values = raw_sensor.read_info(*keys)
+                        if getattr(raw_sensor, "centroid_found", False) or not np.all(
+                            np.isfinite(values)
+                        ):
+                            raise ValueError(
+                                "unloaded reference contains contact or invalid readings"
+                            )
                         zero_samples.append(values)
                         if len(zero_samples) == zero_frames:
-                            raw = {axis: float(np.median([value[index] for value in zero_samples]))
-                                   for index, axis in enumerate(("fn", "fx", "fy")) if axis in model.channels}
-                            sensor.set_calibration_reference(raw, reference_id=f"video-unloaded-{row.frame_id}")
+                            raw = {
+                                axis: float(
+                                    np.median([value[index] for value in zero_samples])
+                                )
+                                for index, axis in enumerate(("fn", "fx", "fy"))
+                                if axis in raw_sensor.calibration.channels
+                            }
+                            raw_sensor.set_calibration_reference(
+                                raw,
+                                reference_id=f"video-unloaded-{row.frame_id}",
+                            )
                             provenance.counts["unloaded_reference_frames"] = zero_frames
                             provenance.counts["unloaded_reference_fn_raw"] = raw.get("fn")
                     force = self._read_force(chain)
@@ -945,8 +1221,16 @@ class AnalysisService:
                         depth=depth_array,
                         force=force,
                     )
-                if zero_frames and len(zero_samples) != zero_frames:
-                    raise ValueError("video ended before 30 unloaded reference frames were available")
+                if field_weighted:
+                    if zero_frames and state.field_baseline is None:
+                        raise ValueError(
+                            "video ended before the 30-frame weighted unloaded baseline "
+                            "was available"
+                        )
+                elif zero_frames and len(zero_samples) != zero_frames:
+                    raise ValueError(
+                        "video ended before 30 unloaded reference frames were available"
+                    )
 
     # ------------------------------------------------------------------- internals
     def _zero_reference_frames(self) -> int:
@@ -1055,6 +1339,7 @@ class AnalysisService:
         self._accumulate(
             provenance.counts, samples, live.depth_nonzero, live.depth_abs_sum
         )
+        field_quality = None if live.force is None else live.force.field_quality()
         record = encode_frame(
             frame_id=row.frame_id,
             t_rel_s=row.t_rel_s,
@@ -1068,6 +1353,7 @@ class AnalysisService:
             depth_nonzero=live.depth_nonzero,
             dropped=row.dropped,
             force=live.force,
+            field_quality=field_quality,
         )
         counts = provenance.counts
         counts["frame_record_bytes"] = counts.get("frame_record_bytes", 0) + len(record)
@@ -1153,6 +1439,13 @@ class AnalysisService:
         counts["force_zero_unavailable_frames"] = counts.get(
             "force_zero_unavailable_frames", 0
         ) + (1 if zero_missing else 0)
+        quality = force.field_quality()
+        if quality is not None:
+            counts["field_quality_frames"] = counts.get("field_quality_frames", 0) + 1
+            counts["field_extrapolated_last"] = quality[0]
+            counts["field_extrapolated_max"] = max(
+                float(counts.get("field_extrapolated_max", 0.0)), quality[0]
+            )
 
     def _provenance(self, key: ResultKey, facts: dict[str, Any]) -> Provenance:
         session = facts["session"]

shared/tests/_field_fakes.py

@@ -0,0 +1,292 @@
+"""Fake SDK surface for the S4 field-model tests (shared by two test modules).
+
+The real SDK field model is exercised by ``orisys-sdk/tests/test_field_calibration.py``;
+the app tests describe the *calls and data* the app owns. Keeping the fakes in one
+place means the pass test and the integration test cannot drift apart.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import math
+from pathlib import Path
+from types import SimpleNamespace
+
+import numpy as np
+
+from orisys_app_shared.acquisition import sdk_worker
+from orisys_app_shared.analysis import service as service_module
+from orisys_app_shared.analysis.model import Provenance
+from orisys_app_shared.record.playback import RecordingFrame
+
+FIELD_PROFILE = "hhd_normal_l1_field/1"
+
+
+# ------------------------------------------------------------------ calibration
+def write_model_file(tmp_path: Path, model_type: str) -> Path:
+    """A schema 3 calibration document the SDK accepts, with its coordinate map.
+
+    ``field_gain/1`` and ``spatial_gain/1`` differ only in the model block, so a test
+    can compare two documents that really load through the SDK's parser.
+    """
+    from orisys.field_calibration import BASELINE_POLICY, FIELD_IDENTITY, MAPPING_POLICY
+
+    rows, cols = np.mgrid[0:4, 0:4]
+    uv = np.stack((cols, rows), axis=-1).astype(np.float64)
+    attachment = tmp_path / "map.npz"
+    np.savez(
+        attachment,
+        uv_by_output_px=uv,
+        valid_mask=np.ones((4, 4), dtype=bool),
+        origin_xy=np.array([0, 0]),
+    )
+    digest = hashlib.sha256(attachment.read_bytes()).hexdigest()
+    pipeline = {
+        "stitch_resolution": [4, 4],
+        "dis_mode": 1,
+        "hhd_downsample": 1,
+        "filters": {"measurement": {"flow_ema_alpha": 1.0}},
+        "feature_signature_sha256": "a" * 64,
+        "geometry_sha256": "b" * 64,
+        "source_scale": 0.42,
+        "coordinates": {
+            "frame": "autocal_flat",
+            "unit": "px",
+            "origin_xy": [0, 0],
+            "map_sha256": digest,
+        },
+    }
+    params = {
+        "points_uv": [[0, 0], [3, 0], [0, 3], [3, 3]],
+        "gains_N_per_raw": [1.0, 1.0, 1.0, 1.0],
+        "triangles": [[0, 1, 2], [1, 3, 2]],
+        "valid_force_range_N": [0.0, 3.0],
+    }
+    if model_type == "field_gain/1":
+        pipeline["field"] = dict(FIELD_IDENTITY)
+        profile = FIELD_PROFILE
+        params["mapping"] = dict(MAPPING_POLICY)
+        params["baseline"] = dict(BASELINE_POLICY)
+    else:
+        profile = "fn_delta_autocal_uv/1"
+    document = {
+        "schema_version": 3,
+        "unit": "N",
+        "identity": {"sensor_type": "finger", "sn": "dd02-gl"},
+        "pipeline": pipeline,
+        "model": {
+            "type": model_type,
+            "input_profile": profile,
+            "output_frame": "sdk_force/1",
+            "outputs": ["fn"],
+            "params": params,
+            "artifacts": [
+                {"role": "coordinate_map", "path": attachment.name, "sha256": digest}
+            ],
+        },
+    }
+    path = tmp_path / ("field.json" if model_type == "field_gain/1" else "spatial.json")
+    path.write_text(json.dumps(document), encoding="utf-8")
+    return path
+
+
+# --------------------------------------------------------------------- recording
+def minimal_recording(root: Path, *, segments: int = 1) -> Path:
+    """The two text files and the segment bytes a reader needs, and nothing else."""
+    recording = root / "recording"
+    recording.mkdir(parents=True, exist_ok=True)
+    (recording / "frames.csv").write_text(
+        "frame_id,t_rel_s,acquired_at,segment,segment_frame,in_video,dropped\n"
+        "1,0.0,,1,1,1,0\n2,0.033,,1,2,1,0\n",
+        encoding="utf-8",
+    )
+    for number in range(1, segments + 1):
+        (recording / f"raw_{number:04d}.mp4").write_bytes(b"")
+    (recording / "session.json").write_text(
+        json.dumps(
+            {
+                "session_id": 1,
+                "started_at": "2026-01-01T00:00:00+0000",
+                "ended_at": "",
+                "device": {"sensor_type": "finger", "config_ref": ""},
+                "stats": {"frames": 2, "video_frames": 2, "segments": segments},
+            }
+        ),
+        encoding="utf-8",
+    )
+    return recording
+
+
+def field_rows(count: int, *, segment: int = 1) -> list[RecordingFrame]:
+    return [
+        RecordingFrame(i + 1, i / 30.0, "", segment, i + 1, True)
+        for i in range(count)
+    ]
+
+
+# ------------------------------------------------------------------------- SDK
+class FakeFieldModel:
+    schema_version = 3
+    input_profile = FIELD_PROFILE
+    model_type = "field_gain/1"
+    channels = {"fn": "N"}
+    fingerprint = "abcdef0123456789" * 4
+
+    def __init__(self) -> None:
+        self.field_baseline_N: float | None = None
+
+
+class FakeFieldSensor:
+    """The slice of ``Sensor`` the pass touches, simulating the SDK's window."""
+
+    def __init__(self, *, zero_frames: int = 30) -> None:
+        self.info = SimpleNamespace(
+            FNORMAL=0,
+            FSHEARX=1,
+            FSHEARY=2,
+            FNORMAL_RAW=3,
+            FSHEARX_RAW=4,
+            FSHEARY_RAW=5,
+        )
+        self.calibration = FakeFieldModel()
+        self.force_calibrated = True
+        self.centroid_found = True  # the field model must not gate on it
+        self.flow = np.zeros((4, 4, 2), dtype=np.float32)
+        self.valid_mask = None
+        self.calibration_status = {"fn": "zero_unavailable"}
+        self.calibration_result = None
+        self.declared: list[str] = []
+        self.set_reference_calls: list[tuple] = []
+        self.pending = 0
+        self.frame_index = 0
+        self.fn_value = float("nan")
+        self.zero_frames = zero_frames
+        self._hhd_downsample = 2
+        self._hhd_grid_shape = (2, 2)
+        self._hhd_ref_shape = (4, 4)
+
+    def declare_unloaded_reference(self, *, reference_id: str) -> None:
+        self.declared.append(str(reference_id))
+        self.pending = 0
+        self.calibration.field_baseline_N = None
+
+    def set_calibration_reference(self, raw, *, reference_id) -> None:  # pragma: no cover
+        self.set_reference_calls.append((raw, reference_id))
+
+    def read_info(self, *keys):
+        values = {
+            0: self.fn_value,
+            1: float("nan"),
+            2: float("nan"),
+            3: 100.0,
+            4: 2.0,
+            5: 3.0,
+        }
+        return tuple(values[key] for key in keys)
+
+    def step(self):
+        """One frame of the SDK's field model: 30-frame window, then ok frames."""
+        self.frame_index += 1
+        diagnostics: dict = {}
+        if self.declared:
+            if self.pending < self.zero_frames:
+                self.pending += 1
+                diagnostics = {
+                    "pending_reference_frames": self.pending,
+                    "extrapolated_fraction": 0.4,
+                    "coverage_fraction": 0.9,
+                    "model_fingerprint": self.calibration.fingerprint,
+                    "frame_id": self.frame_index,
+                }
+                if self.pending == self.zero_frames:
+                    self.calibration.field_baseline_N = 0.25
+                    diagnostics["baseline_N"] = 0.25
+                    diagnostics["pending_reference_frames"] = 0
+                    diagnostics["extrapolated_fraction"] = 0.3
+                self.calibration_status = {"fn": "zero_unavailable"}
+                self.fn_value = float("nan")
+            else:
+                diagnostics = {
+                    "baseline_N": 0.25,
+                    "extrapolated_fraction": 0.2,
+                    "coverage_fraction": 0.97,
+                    "model_fingerprint": self.calibration.fingerprint,
+                }
+                self.calibration_status = {"fn": "ok"}
+                self.fn_value = 1.5
+        else:
+            diagnostics = {"model_fingerprint": self.calibration.fingerprint}
+        self.calibration_result = SimpleNamespace(
+            diagnostics=diagnostics,
+            status=dict(self.calibration_status),
+            values_N={"fn": None if math.isnan(self.fn_value) else self.fn_value},
+        )
+        return None
+
+
+class FakeChain:
+    def __init__(self, sensor, *, frames: int = 35) -> None:
+        self.sensor = sensor
+        self.frames = int(frames)
+        self.index = 0
+        self.open_seconds = 0.0
+        self.debug_files: list[str] = []
+        self.decode_position = 0
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *args):
+        return False
+
+    def read(self):
+        if self.index >= self.frames:
+            return False, None
+        self.index += 1
+        return True, np.zeros((4, 4), dtype=np.uint8)
+
+    def feed(self, frame):
+        return frame
+
+
+class FakeReader:
+    def __init__(self, rows, *, segments: int = 1, base: Path) -> None:
+        self._rows = list(rows)
+        self._segments = [
+            base / f"raw_{number:04d}.mp4" for number in range(1, segments + 1)
+        ]
+
+    def session(self):
+        return {
+            "session_id": 1,
+            "started_at": "2026-01-01T00:00:00+0000",
+            "ended_at": "",
+            "device": {"sensor_type": "finger", "config_ref": ""},
+            "stats": {"frames": len(self._rows), "segments": len(self._segments)},
+        }
+
+    def info(self):
+        return SimpleNamespace(frames=self._rows, video_frames=len(self._rows))
+
+    def segments(self):
+        return list(self._segments)
+
+    def segment_path(self, number: int):
+        index = int(number) - 1
+        return self._segments[index] if 0 <= index < len(self._segments) else None
+
+
+def install_pass(monkeypatch, processor, chain: FakeChain, reader: FakeReader) -> None:
+    """Make the service walk the fake chain instead of opening the SDK."""
+    monkeypatch.setattr(processor, "reader", lambda: reader)
+    monkeypatch.setattr(processor, "_segment_offset", lambda *args: (0, True))
+    monkeypatch.setattr(service_module, "_OpenChain", lambda *args: chain)
+    monkeypatch.setattr(sdk_worker, "process_chain", lambda sensor: (sensor.step(), ""))
+
+
+def new_pass_state(processor, tmp_path: Path, *, count: int = 35):
+    rows = field_rows(count)
+    return service_module._PassState(
+        Provenance(), [1], {1: list(rows)}, tmp_path
+    )

shared/tests/test_analysis_force_validity.py

@@ -10,6 +10,8 @@ from __future__ import annotations
 
 import math
 
+import pytest
+
 from orisys_app_shared.analysis.model import ALGORITHM_VERSION, ForceSample
 from orisys_app_shared.analysis.service import AnalysisService
 
@@ -97,3 +99,45 @@ def test_results_stored_under_the_old_validity_rule_are_not_served():
     # Version 1 marked every frame invalid when the model lacked a zero; the cache key
     # carries this string, so those results must miss the cache and be recomputed.
     assert ALGORITHM_VERSION != "flow-hhd-contact-1"
+
+
+def test_extrapolated_fraction_is_quality_not_validity():
+    """A field estimate's extrapolation is evidence about its support, not validity.
+
+    The SDK reports the field model's extrapolated/coverage fraction in
+    ``calibration_diagnostics`` and keeps the per-axis status ``ok``; the app stores
+    the fractions as quality and must keep the newton value. Reading them as
+    invalidity would silently drop every experimental value.
+    """
+    reading = _reading({"fn": "ok"}, normal=3.5)
+    reading["calibration_diagnostics"] = {
+        "extrapolated_fraction": 0.42,
+        "coverage_fraction": 0.9,
+        "baseline_N": 0.25,
+        "model_fingerprint": "f" * 64,
+    }
+    sample = ForceSample.from_reading(reading)
+    assert sample.normal == pytest.approx(3.5)
+    assert sample.valid
+    assert sample.diagnostics["extrapolated_fraction"] == pytest.approx(0.42)
+    assert sample.field_quality() == pytest.approx((0.42, 0.9))
+
+    # The force payload is unchanged by the new evidence: quality is stored in its
+    # own frame payload, so an older result and this one read the same force fields.
+    decoded = ForceSample.decode(sample.encode())
+    assert decoded.normal == pytest.approx(3.5)
+    assert decoded.diagnostics == {}
+
+
+def test_a_baseline_wait_still_keeps_the_raw_measurement():
+    """The first 30 declared frames are ``zero_unavailable``: raw, not newton."""
+    reading = _reading({"fn": "zero_unavailable"}, normal=float("nan"))
+    reading["calibration_diagnostics"] = {
+        "pending_reference_frames": 7,
+        "extrapolated_fraction": 0.35,
+    }
+    sample = ForceSample.from_reading(reading)
+    assert sample.valid
+    assert sample.raw() == (1234.0, -5.0, 7.0)
+    assert math.isnan(sample.normal)
+    assert sample.diagnostics["pending_reference_frames"] == 7

shared/tests/test_field_calibration_app.py

@@ -0,0 +1,339 @@
+"""S4: the field-weighted model in the replay pass, result payload and live entries.
+
+The field model is an *experiment*: the SDK computes ``P = A * sum(g * q)`` on the
+HHD grid and subtracts a baseline measured from a declared 30-frame unloaded
+window. The app's part is to declare that window exactly once, keep the numeric
+quality (extrapolated fraction) out of the validity decision, store it per frame,
+and refuse the model at the live entries it was not wired into.
+
+The SDK is faked in ``_field_fakes``: these tests describe the app-side contract
+(which calls it makes, in which order, and what it stores), not the SDK's own
+arithmetic, which ``orisys-sdk/tests/test_field_calibration.py`` owns.
+"""
+
+from __future__ import annotations
+
+import math
+from types import SimpleNamespace
+
+import numpy as np
+import pytest
+
+from _field_fakes import (
+    FakeChain,
+    FakeFieldSensor,
+    FakeReader,
+    field_rows,
+    install_pass,
+    minimal_recording,
+    new_pass_state,
+    write_model_file,
+)
+from orisys_app_shared.acquisition.finger_worker import FingerAcquisitionWorker
+from orisys_app_shared.acquisition.sdk_worker import (
+    SdkAcquisitionWorker,
+    live_field_model_rejection,
+    read_forces_from_sensor,
+)
+from orisys_app_shared.analysis import model as model_module
+from orisys_app_shared.analysis.cache import AnalysisWriter, ResultCache
+from orisys_app_shared.analysis.model import (
+    FIELD_KEY_SUFFIX,
+    FIELD_PAYLOAD,
+    FIELD_PROFILE,
+    FRAME_FLAG_FIELD,
+    ForceSample,
+    Provenance,
+    decode_frame,
+    encode_frame,
+    frame_payload_bytes,
+    result_key,
+)
+from orisys_app_shared.analysis.reader import AnalysisReader
+from orisys_app_shared.analysis.service import (
+    AnalysisService,
+    ChainSpec,
+    calibration_profile,
+)
+from orisys_app_shared.record.playback import RecordingFrame
+
+
+# ------------------------------------------------------------------- payload
+def test_field_quality_round_trips_in_the_frame_record():
+    payload = {
+        "frame_id": 7,
+        "t_rel_s": 0.25,
+        "grid_id": 0,
+        "dx": np.zeros(0, dtype=np.float32),
+        "dy": np.zeros(0, dtype=np.float32),
+        "magnitude": np.zeros(0, dtype=np.float32),
+        "active": np.zeros(0, dtype=bool),
+        "field_quality": (0.25, 0.9),
+    }
+    decoded = decode_frame(encode_frame(**payload))
+    assert decoded.field_quality == pytest.approx((0.25, 0.9))
+    assert decoded.frame_id == 7 and decoded.t_rel_s == pytest.approx(0.25)
+
+    legacy = decode_frame(encode_frame(**{**payload, "field_quality": None}))
+    assert legacy.field_quality is None, "a frame without the flag must read as unknown"
+
+    assert frame_payload_bytes(0) == frame_payload_bytes(0, with_field_quality=False)
+    assert (
+        frame_payload_bytes(0, with_field_quality=True)
+        == frame_payload_bytes(0) + FIELD_PAYLOAD.size
+    )
+    assert FRAME_FLAG_FIELD != 0 and FIELD_PROFILE == "hhd_normal_l1_field/1"
+
+
+def test_a_frame_with_force_and_field_quality_decodes_both():
+    sample = ForceSample.from_reading(
+        {
+            "normal_force": 1.5,
+            "normal_force_raw": 100.0,
+            "valid": True,
+            "force_calibrated": True,
+            "calibrated_axes": ("fn",),
+            "calibration_status": {"fn": "ok"},
+            "calibration_diagnostics": {
+                "extrapolated_fraction": 0.3,
+                "coverage_fraction": 0.9,
+            },
+        }
+    )
+    record = encode_frame(
+        frame_id=3,
+        t_rel_s=0.0,
+        grid_id=0,
+        dx=np.zeros(0, dtype=np.float32),
+        dy=np.zeros(0, dtype=np.float32),
+        magnitude=np.zeros(0, dtype=np.float32),
+        active=np.zeros(0, dtype=bool),
+        force=sample,
+        field_quality=sample.field_quality(),
+    )
+    decoded = decode_frame(record)
+    assert decoded.force is not None and decoded.force.normal == pytest.approx(1.5)
+    assert decoded.field_quality == pytest.approx((0.3, 0.9))
+    # Quality is evidence about the estimate, not part of the force payload.
+    assert decoded.force.diagnostics == {}
+
+
+def test_the_reader_serves_field_quality_and_none_for_old_frames(tmp_path):
+    key = result_key(recording="r", sensor_config="s", calibration="c")
+    directory = tmp_path / "result"
+    writer = AnalysisWriter(directory, key=key)
+    writer.begin()
+    grid = writer.grid_keyframe(np.zeros(0, np.int32), np.zeros(0, np.int32))
+    for frame_id, quality in ((1, (0.4, 0.8)), (2, None)):
+        writer.add_frame(
+            encode_frame(
+                frame_id=frame_id,
+                t_rel_s=frame_id / 30.0,
+                grid_id=grid,
+                dx=np.zeros(0, dtype=np.float32),
+                dy=np.zeros(0, dtype=np.float32),
+                magnitude=np.zeros(0, dtype=np.float32),
+                active=np.zeros(0, dtype=bool),
+                field_quality=quality,
+            ),
+            frame_id=frame_id,
+            t_rel_s=frame_id / 30.0,
+        )
+    writer.commit(provenance=Provenance(), counts={})
+
+    reader = AnalysisReader(directory)
+    try:
+        first = reader.frame(1)
+        second = reader.frame(2)
+        assert first.field_quality == pytest.approx((0.4, 0.8))
+        assert second.field_quality is None
+    finally:
+        reader.close()
+
+
+# ----------------------------------------------------------------------- key
+def test_field_calibration_adds_its_own_cache_component(tmp_path):
+    recording = minimal_recording(tmp_path)
+    field_file = write_model_file(tmp_path, "field_gain/1")
+    spatial_file = write_model_file(tmp_path, "spatial_gain/1")
+
+    assert calibration_profile(field_file) == ("field_gain/1", FIELD_PROFILE)
+    assert calibration_profile(spatial_file)[1] == "fn_delta_autocal_uv/1"
+    assert calibration_profile(None) == ("", "")
+
+    field_service = AnalysisService(
+        recording, chain=ChainSpec("finger", calibration=field_file)
+    )
+    spatial_service = AnalysisService(
+        recording, chain=ChainSpec("finger", calibration=spatial_file)
+    )
+    field_key = field_service.dependency_report()["key"]
+    spatial_key = spatial_service.dependency_report()["key"]
+
+    assert field_key.components["recording"] == spatial_key.components["recording"]
+    assert field_key.components["sensor_config"].endswith(FIELD_KEY_SUFFIX)
+    assert FIELD_KEY_SUFFIX not in spatial_key.components["sensor_config"]
+    assert field_service.dependency_report()["model_profile"] == FIELD_PROFILE
+    assert field_key.digest != spatial_key.digest
+
+
+# ---------------------------------------------------------------------- pass
+def test_field_pass_declares_one_window_and_does_not_use_contact_or_scalar_zero(
+    monkeypatch, tmp_path
+):
+    sensor = FakeFieldSensor()
+    chain = FakeChain(sensor, frames=35)
+    recording = minimal_recording(tmp_path)
+    reader = FakeReader(field_rows(35), base=recording)
+    processor = AnalysisService(
+        recording,
+        chain=ChainSpec("finger"),
+        store_force=True,
+        unloaded_reference_frames=30,
+    )
+    install_pass(monkeypatch, processor, chain, reader)
+
+    state = new_pass_state(processor, tmp_path)
+    frames = list(processor._pass_frames(state))
+
+    assert len(frames) == 35
+    assert sensor.declared == ["video-unloaded-1"], "the window is declared exactly once"
+    assert sensor.set_reference_calls == [], "field models must not get a scalar raw zero"
+    # 30 pending frames are zero_unavailable; their raw counts stay the measurement.
+    for live in frames[:30]:
+        assert dict(live.force.status)["fn"] == "zero_unavailable"
+        assert live.force.valid and math.isnan(live.force.normal)
+        assert live.force.normal_raw == pytest.approx(100.0)
+    assert dict(frames[30].force.status)["fn"] == "ok"
+    assert frames[30].force.normal == pytest.approx(1.5)
+    # The SDK's weighted baseline, not a median of raw counts.
+    assert state.field_baseline == pytest.approx(0.25)
+    assert state.field_reference_id == "video-unloaded-1"
+    assert state.field_extrapolated_count == 35
+    # centroid_found was True throughout: the field model has no contact gate.
+
+
+def test_field_pass_records_provenance_and_per_frame_quality(monkeypatch, tmp_path):
+    sensor = FakeFieldSensor()
+    chain = FakeChain(sensor, frames=35)
+    recording = minimal_recording(tmp_path)
+    reader = FakeReader(field_rows(35), base=recording)
+    field_file = write_model_file(tmp_path, "field_gain/1")
+    cache = ResultCache(tmp_path / "cache")
+    processor = AnalysisService(
+        recording,
+        chain=ChainSpec("finger", calibration=field_file),
+        cache=cache,
+        store_force=True,
+        unloaded_reference_frames=30,
+    )
+    install_pass(monkeypatch, processor, chain, reader)
+
+    outcome = processor.run()
+    assert outcome.complete, outcome.reason
+
+    reader_out = AnalysisReader(outcome.directory)
+    try:
+        provenance = reader_out.provenance
+        field = dict(provenance.field)
+        assert field["profile"] == FIELD_PROFILE
+        assert field["strategy"] == model_module.FIELD_STRATEGY
+        assert field["baseline_N"] == pytest.approx(0.25)
+        assert field["reference_id"] == "video-unloaded-1"
+        assert field["frames"] == 35
+        expected_mean = (29 * 0.4 + 0.3 + 5 * 0.2) / 35
+        assert field["extrapolated_mean"] == pytest.approx(expected_mean, abs=1e-6)
+        assert field["extrapolated_max"] == pytest.approx(0.4)
+        assert field["hhd"]["downsample"] == 2
+        assert field["hhd"]["grid_shape"] == [2, 2]
+        assert reader_out.counts["field_quality_frames"] == 35
+        # The same window's frames carry quality in the record; frame ids are the
+        # recording's own, not the SDK's diagnostic id.
+        first = reader_out.frame(1)
+        last = reader_out.frame(35)
+        assert first.field_quality == pytest.approx((0.4, 0.9))
+        assert last.field_quality == pytest.approx((0.2, 0.97))
+        assert reader_out.key_components["sensor_config"].endswith(FIELD_KEY_SUFFIX)
+    finally:
+        reader_out.close()
+
+
+def test_field_model_rejects_a_multi_segment_recording(monkeypatch, tmp_path):
+    field_file = write_model_file(tmp_path, "field_gain/1")
+    recording = minimal_recording(tmp_path, segments=2)
+    rows = field_rows(2) + [
+        RecordingFrame(i + 1, (i + 2) / 30.0, "", 2, i + 1, True) for i in range(2)
+    ]
+    sensor = FakeFieldSensor()
+    chain = FakeChain(sensor, frames=4)
+    reader = FakeReader(rows, segments=2, base=recording)
+    cache = ResultCache(tmp_path / "cache")
+    processor = AnalysisService(
+        recording,
+        chain=ChainSpec("finger", calibration=field_file),
+        cache=cache,
+        store_force=True,
+        unloaded_reference_frames=30,
+    )
+    install_pass(monkeypatch, processor, chain, reader)
+
+    with pytest.raises(ValueError, match="single-segment"):
+        processor.run()
+    assert not list((tmp_path / "cache").glob("*/meta.json"))
+
+
+# --------------------------------------------------------------------- live
+def test_live_entries_refuse_a_field_model_file(tmp_path):
+    field_file = write_model_file(tmp_path, "field_gain/1")
+    rejection = live_field_model_rejection(field_file)
+    assert FIELD_PROFILE in rejection and "replay" in rejection.lower()
+
+    box = SdkAcquisitionWorker()
+    with pytest.raises(RuntimeError, match="hhd_normal_l1_field/1"):
+        box.start(0, config_ref=tmp_path / "unused.json", cal_path=field_file)
+
+    finger = FingerAcquisitionWorker()
+    with pytest.raises(RuntimeError, match="hhd_normal_l1_field/1"):
+        finger.start(0, config_ref=tmp_path / "unused.json", cal_path=field_file)
+
+    # A hot swap is refused before the device is touched too (the sensor double only
+    # has to exist: the refusal comes before any SDK call on it).
+    box._pending_calibration = str(field_file)
+    with pytest.raises(RuntimeError, match="hhd_normal_l1_field/1"):
+        box._consume_pending_calibration(sensor=SimpleNamespace())
+    finger._pending_calibration = str(field_file)
+    with pytest.raises(RuntimeError, match="hhd_normal_l1_field/1"):
+        finger._consume_pending_calibration(sensor=SimpleNamespace())
+
+
+def test_a_scalar_calibration_is_still_accepted_by_the_live_entry(tmp_path):
+    spatial_file = write_model_file(tmp_path, "spatial_gain/1")
+    assert live_field_model_rejection(spatial_file) == ""
+    assert live_field_model_rejection(None) == ""
+    assert live_field_model_rejection(tmp_path / "missing.json") == ""
+
+
+def test_the_field_reading_carries_quality_without_touching_validity():
+    sensor = FakeFieldSensor()
+    sensor.declare_unloaded_reference(reference_id="unloaded-1")
+    sensor.step()
+    reading = read_forces_from_sensor(sensor)
+    assert reading["calibration_status"] == {"fn": "zero_unavailable"}
+    assert reading["calibration_diagnostics"]["pending_reference_frames"] == 1
+    sample = ForceSample.from_reading(reading)
+    assert sample.valid, "a missing baseline is not an invalid raw measurement"
+    assert math.isnan(sample.normal) and sample.normal_raw == pytest.approx(100.0)
+    # Quality is available while the baseline is pending: it describes the support of
+    # this frame's integral, not whether the baseline is ready.
+    assert sample.field_quality() == pytest.approx((0.4, 0.9))
+    for _ in range(30):
+        sensor.step()
+    ready = read_forces_from_sensor(sensor)
+    assert ready["calibration_status"] == {"fn": "ok"}
+    ready_sample = ForceSample.from_reading(ready)
+    assert ready_sample.normal == pytest.approx(1.5)
+    assert ready_sample.field_quality() == pytest.approx((0.2, 0.97))
+
+    # A reading with no diagnostics at all is *unknown*, never 0 extrapolation.
+    reading.pop("calibration_diagnostics")
+    assert ForceSample.from_reading(reading).field_quality() is None

shared/tests/test_multical_integration.py

@@ -182,3 +182,39 @@ def test_video_unloaded_reference_is_explicit_and_requires_the_whole_interval(mo
         assert len(list(processor._pass_frames(state))) == frame_count
         assert applied == [{"fn": 100.}]
         assert state.provenance.counts["unloaded_reference_frames"] == 30
+
+
+def test_a_field_model_reading_keeps_raw_and_reports_weighted_quality(tmp_path):
+    """The field model is a different zeroing protocol, not a different validity rule.
+
+    Its reading carries the SDK's weighted diagnostics; the app must keep the raw
+    counts while the baseline is pending and report the extrapolated fraction as
+    quality (0.0 is a real value, "no diagnostics" is unknown).
+    """
+    from orisys_app_shared.acquisition.sdk_worker import read_forces_from_sensor
+    from orisys_app_shared.analysis.model import ForceSample
+
+    from _field_fakes import FakeFieldSensor  # tests' own SDK double
+
+    sensor = FakeFieldSensor()
+    sensor.declare_unloaded_reference(reference_id="unloaded-1")
+    sensor.step()
+    reading = read_forces_from_sensor(sensor)
+    assert reading["calibration_status"] == {"fn": "zero_unavailable"}
+    assert reading["calibration_diagnostics"]["pending_reference_frames"] == 1
+    sample = ForceSample.from_reading(reading)
+    assert sample.valid and sample.normal_raw == pytest.approx(100.0)
+    assert math.isnan(sample.normal)
+    # Quality is a statement about the frame's support, available while the baseline
+    # is still pending; only absent diagnostics mean "unknown".
+    assert sample.field_quality() == pytest.approx((0.4, 0.9))
+    reading.pop("calibration_diagnostics")
+    assert ForceSample.from_reading(reading).field_quality() is None
+
+    for _ in range(30):
+        sensor.step()
+    ready = read_forces_from_sensor(sensor)
+    assert ready["calibration_status"] == {"fn": "ok"}
+    ready_sample = ForceSample.from_reading(ready)
+    assert ready_sample.normal == pytest.approx(1.5)
+    assert ready_sample.field_quality() == pytest.approx((0.2, 0.97))