fix: harden v1.2 trajectory disposal matching

.gitignore

@@ -3,6 +3,7 @@ __pycache__/
 .DS_Store
 *.textClipping
 .pytest_cache/
+.superpowers/
 .venv/
 dist/
 build/

README_zh.md

@@ -20,7 +20,9 @@
 
 ## 当前实现范围
 
-当前版本先实现纯业务状态机，不依赖摄像头模型。后续视觉模块只需要输出标准观察数据：
+当前版本已经接入可运行的轻量视觉流程：区域占用、垃圾桶动作和 v1.2 的轻量 motion trajectory 都使用启发式图像差分实现，不使用 YOLO。后续训练好的 YOLO 食品检测模型会通过统一的 `disposal_evidence` / backend 合约接入，不改变批次计时状态机的业务输入形态。
+
+视觉或 backend 模块需要输出标准观察数据：
 
 ```json
 {
@@ -29,7 +31,23 @@
     "1": 1,
     "2": 0
   },
-  "trash_deposit": false
+  "trash_deposit": false,
+  "disposal_evidence": [
+    {
+      "source_zone_id": "1",
+      "target": "trash",
+      "confidence": 0.9,
+      "method": "motion",
+      "track_points": [
+        {"x": 0.22, "y": 0.30},
+        {"x": 0.48, "y": 0.58},
+        {"x": 0.76, "y": 0.78}
+      ],
+      "item_class": null,
+      "detector_score": null,
+      "observed_at": "2026-04-27T10:00:03+08:00"
+    }
+  ]
 }
 ```
 
@@ -132,15 +150,17 @@ scripts/run_runtime.sh
 2. 用 `ffmpeg` 周期抓取小尺寸 RGB 帧。
 3. 按标定区域做占用变化检测。
 4. 判断垃圾桶区域是否有明显投放动作。
-5. 调用批次计时状态机。
-6. 写入 `logs/events.jsonl`，管理页会读取这个文件。
+5. 对刚清空的来源区域运行轻量 motion trajectory，生成可选的 `disposal_evidence`。
+6. 调用批次计时状态机，优先使用匹配 `source_zone_id` 的 `disposal_evidence` 确认丢弃，再回退到通用垃圾桶动作。
+7. 写入 `logs/events.jsonl`，管理页会读取这个文件。
 
 当前视觉版本是可运行的启发式版本：
 
 - 每个格口输出 `0/1` 占用状态，不识别单份数量。
 - 启动后的前几帧用于建立空柜基线，默认 `3` 帧。
 - 如果启动时格口里已经有食品，系统会把它当作基线，后续要等画面变化后才会产生计时事件。
-- 真实生产精度后续应接食品检测模型。
+- v1.2 轨迹识别是轻量 motion trajectory，不加载 YOLO，不要求模型文件。
+- 训练好的 YOLO 模型后续应作为新的 backend 接入，并继续输出统一的 `disposal_evidence`。
 
 可选运行参数可以放在配置文件的 `[runtime]` 中：
 
@@ -167,9 +187,29 @@ trash_motion_delta = 18.0
 trash_sustained_motion_delta = 8.0
 trash_sustained_motion_frames = 2
 trash_motion_cooldown_seconds = 3
+trajectory_enabled = true
+trajectory_window_seconds = 8
+trajectory_sample_interval_seconds = 1.0
+trajectory_min_points = 3
+trajectory_min_confidence = 0.72
+trajectory_motion_delta = 20.0
+trajectory_min_blob_area = 12
+trajectory_max_blob_area_fraction = 0.35
+trajectory_trash_entry_margin = 0.04
+trajectory_backend = "motion"
+yolo_enabled = false
+yolo_model_path = ""
+yolo_min_confidence = 0.65
 diagnostics_path = "logs/runtime_diagnostics.jsonl"
 ```
 
+`trajectory_backend = "motion"` 表示当前使用轻量轨迹 backend。`yolo_enabled`、`yolo_model_path` 和 `yolo_min_confidence` 是为后续训练模型预留的配置项；当前版本即使保留这些字段，也不会启用 YOLO 推理。
+
+运行诊断写入 `logs/runtime_diagnostics.jsonl`。每行包含顶层 `disposal_evidence`，以及 `diagnostics.trajectory`：
+
+- 顶层 `disposal_evidence`：本帧实际输出给状态机的来源区域到垃圾桶证据。
+- `diagnostics.trajectory`：轻量轨迹 backend 的候选、过期、拒绝和已发出证据等调试信息。
+
 ## 本地测试
 
 ```bash

docs/project.md

@@ -17,7 +17,7 @@ The `v1.2 轨迹识别` batch adds source-zone trajectory evidence for disposal
   - `manage_api.py`: standard-library HTTP management API.
   - `main.py`: RTSP runtime loop connecting frame capture, vision, state engine, and JSONL sinks.
   - `vision.py`: heuristic ROI occupancy and trash-motion detection.
-  - v1.2 adds trajectory evidence between vision and engine: `MotionTrajectoryBackend` emits source-zone-to-trash evidence; `BatchEngine` consumes the backend-neutral `disposal_evidence` contract.
+  - v1.2 adds trajectory evidence between vision and engine: `TrajectoryTracker` emits source-zone-to-trash evidence; `BatchEngine` consumes the backend-neutral `disposal_evidence` contract.
 - Frontend package: `web/`
   - Vite + vanilla JavaScript management console.
   - Default web port `23000`.
@@ -26,6 +26,7 @@ The `v1.2 轨迹识别` batch adds source-zone trajectory evidence for disposal
 - Runtime data:
   - Events JSONL default path `logs/events.jsonl`.
   - Diagnostics JSONL default path `logs/runtime_diagnostics.jsonl`.
+  - v1.2 diagnostics include root-level `disposal_evidence` plus `diagnostics.trajectory`.
 - Deployment:
   - Root Python Docker image for API/runtime.
   - `web/Dockerfile` for static web console.
@@ -52,9 +53,34 @@ The `v1.2 轨迹识别` batch adds source-zone trajectory evidence for disposal
   - `trajectory_enabled`: enables source-zone trajectory evidence.
   - `trajectory_window_seconds`: seconds after a zone clears where movement can confirm disposal.
   - `trajectory_sample_interval_seconds`: faster runtime delay while a candidate is active.
+  - `trajectory_min_points`: minimum sampled motion points required before evidence can emit.
   - `trajectory_min_confidence`: minimum confidence before evidence can close pending disposal.
+  - `trajectory_motion_delta`: frame-difference threshold for trajectory motion points.
+  - `trajectory_min_blob_area`: minimum connected motion area to keep as a point.
+  - `trajectory_max_blob_area_fraction`: rejects overly broad frame motion as ambiguous.
+  - `trajectory_trash_entry_margin`: margin for treating a track point as entering the trash ROI.
   - `trajectory_backend`: first valid value is `"motion"`.
-  - `yolo_enabled`: defaults to `false`; reserved for a future trained model backend.
+  - `yolo_enabled`, `yolo_model_path`, `yolo_min_confidence`: reserved for a future trained model backend. Current v1.2 keeps YOLO disabled.
+
+## v1.2 Runtime Flow
+
+1. `RTSPFrameSource` captures a resized RGB frame.
+2. `ZoneOccupancyDetector` updates per-zone binary occupancy and generic trash-motion count from calibrated ROIs.
+3. `TrajectoryTracker` watches zones that just cleared, follows lightweight motion points toward the trash ROI, and emits source-specific `DisposalEvidence` when confidence passes the configured threshold.
+4. `BatchEngine` processes `Observation(zone_counts, trash_deposit_count, disposal_evidence)`.
+5. For pending disposal, matching `disposal_evidence.source_zone_id` confirms `batch_discarded` before generic FIFO `trash_deposit_count` fallback is used.
+6. Runtime writes events to `logs/events.jsonl` and diagnostics to `logs/runtime_diagnostics.jsonl`.
+
+The current tracker is a motion backend only. A later trained YOLO detector should plug in as another backend that enriches or replaces the evidence producer while preserving the same `disposal_evidence` contract consumed by the engine.
+
+## Diagnostics
+
+- Runtime diagnostics JSONL records one item per runtime iteration.
+- Root `disposal_evidence` is the exact evidence list passed into the engine for that iteration.
+- `diagnostics.zones` contains occupancy metrics used to derive `zone_counts`.
+- `diagnostics.trash` contains generic trash-motion metrics and cooldown state.
+- `diagnostics.trajectory` contains v1.2 candidate counts, emitted evidence count, motion point count, and per-candidate emitted/rejected/expired records.
+- Capture failures still keep the v1.2 schema with root `disposal_evidence: []` and `diagnostics.trajectory.reason = "frame_capture_failed"`.
 
 ## Event Model
 
@@ -87,11 +113,17 @@ In v1.2, `batch_discarded` can be triggered by zone-scoped `disposal_evidence` b
   - `scripts/run_runtime.sh`
   - One-frame smoke test when camera and `ffmpeg` are available:
     - `PYTHONPATH=src python3 -m cold_display_guard.main --config config/example.toml --once`
+- v1.2 operating notes:
+  - Keep `trajectory_backend = "motion"` and `yolo_enabled = false` unless a trained YOLO backend has been explicitly deployed.
+  - Confirm `logs/runtime_diagnostics.jsonl` contains top-level `disposal_evidence` and `diagnostics.trajectory` before judging trajectory behavior from events alone.
+  - When `TrajectoryTracker` has active candidates, runtime sampling uses `trajectory_sample_interval_seconds`; this can temporarily be faster than the normal `sample_interval_seconds`.
+  - On remote deployments, preserve the remote `config/example.toml` calibration and stream settings when syncing code.
 
 ## Known Risks
 
 - The current vision detector is heuristic and reports binary occupancy, not item counts.
-- v1.2 motion tracking improves disposal matching but can still miss movement if the hand/object path is occluded or sampled too sparsely.
+- v1.2 motion tracking improves disposal matching but can still miss movement if the hand/object path is occluded, too broad, too small, or sampled too sparsely.
+- YOLO config fields are present for compatibility, but no trained YOLO model is part of the current runtime.
 - If food is already present during baseline collection, those regions may be treated as empty baseline until visual changes occur.
 - Changing calibration while the runtime process has active batches can create operational ambiguity; v1.1 should document or enforce a pause/restart expectation.
 - Historical events must keep the zone index at the time of emission so later region reordering does not reinterpret old logs.

findings.md

@@ -77,6 +77,12 @@ Use `阶段 x` only as a workflow stage inside the same `v1.1 优化改造` batc
 - 轨迹诊断必须记录 active、emitted、rejected、expired，方便现场调参。
 - 后续 YOLO 模型只作为 evidence 增强输入，不能修改 `BatchEngine` 的业务语义。
 
+## Final Review Findings
+
+- Evidence and generic trash fallback must share the same count budget: one matched `disposal_evidence` consumes one observed trash deposit signal when `trash_deposit_count` is also present, but extra trash deposits must still fall back to pending batches.
+- A trajectory candidate must not append source-zone-external motion before source motion has been seen. If outside motion appears first and is later followed by source noise, the candidate is rejected with `motion_started_outside_source` instead of producing evidence.
+- Runtime diagnostics on the deployed host should be checked for schema only, not by printing config, because the remote config may contain RTSP credentials and calibration.
+
 ## Backend Planning Notes
 
 - `EngineSettings.zone_ids` should remain config driven; numeric zones are preferred for new configs, but old `r1c1` style IDs should continue loading.

memories.md

@@ -21,4 +21,19 @@
 - Preserve a stable `disposal_evidence` contract for a future trained YOLO product detector.
 - Keep ROI occupancy timing as the source of zone occupied/empty state.
 - Use trajectory evidence before generic trash-motion FIFO fallback.
-- Remote deployment target is `xiaozheng@192.168.5.206:/home/xiaozheng/cold_display_guard`; preserve remote `config/example.toml`.
+- Current runtime writes top-level `disposal_evidence` and nested `diagnostics.trajectory` into runtime diagnostics JSONL.
+
+## v1.2 Completed Facts
+
+- Stage 1 established the backend contract: `Observation.disposal_evidence` normalizes backend-neutral disposal evidence, and the engine can discard a pending batch only when evidence targets `trash`, meets confidence, and matches the pending `source_zone_id`.
+- Stage 2 added the lightweight motion trajectory runtime path: ROI occupancy still drives occupied/empty state, `TrajectoryTracker` emits source-zone-to-trash evidence, and generic trash-motion count remains as a fallback.
+- Stage 3 added diagnostics and tests for runtime evidence propagation, trajectory sampling interval behavior, capture-failure schema, and trajectory/yolo runtime config parsing.
+- Final review fixes: matched evidence now only subtracts the trash fallback budget by the number of batches it actually closed, and trajectory candidates reject outside-before-source motion with `motion_started_outside_source`.
+- Current v1.2 does not use YOLO. `yolo_enabled`, `yolo_model_path`, and `yolo_min_confidence` are reserved for a future trained model backend that should keep emitting the same `disposal_evidence` shape.
+
+## Remote Deployment Notes
+
+- Remote deployment target is `xiaozheng@192.168.5.206:/home/xiaozheng/cold_display_guard`.
+- Preserve the remote `config/example.toml`; it may contain camera, calibration, threshold, and deployment-specific runtime settings that must not be overwritten blindly.
+- When syncing code remotely, verify that runtime diagnostics still show top-level `disposal_evidence` and `diagnostics.trajectory` before evaluating v1.2 trajectory behavior from `logs/events.jsonl`.
+- The latest v1.2 deployment was verified with `cold-display-guard-runtime` and `cold-display-guard-api` up, API health `status=ok`, and diagnostics schema showing `has_disposal_evidence=True` plus `has_trajectory=True`.

progress.md

@@ -211,6 +211,16 @@
 | 2026-05-29 | Phase 3 | Coding Agent | Fixed runtime integration review concerns | Error diagnostics keep trajectory schema; tests no longer use `create=True`; evidence payload helper type narrowed |
 | 2026-05-29 | Phase 3 | Testing Agent | Re-tested runtime integration concerns | Verdict pass; no new issues |
 | 2026-05-29 | Phase 3 | Main Agent | Ran local verification | `tests.test_main tests.test_vision` passed with 26 tests; full Python suite passed with 68 tests; dependency scan had no model/heavy vision matches |
+| 2026-05-29 | Phase 4 | Main Agent | Marked Phase 4 as `in_progress` | Preparing documentation, final verification, remote deployment, and final review |
+| 2026-05-29 | Phase 4 | Main Agent | Synced v1.2 code to `xiaozheng@192.168.5.206` | `rsync` completed while excluding remote `config/example.toml` |
+| 2026-05-29 | Phase 4 | Main Agent | Tried remote config patch with inline heredoc | Failed with shell quoting `SyntaxError`; switching to scp temporary patch script |
+| 2026-05-29 | Phase 4 | Main Agent | Patched remote runtime config via uploaded Python script | Added missing trajectory/yolo runtime keys without printing RTSP config |
+| 2026-05-29 | Phase 4 | Main Agent | Rebuilt and restarted remote runtime/API containers | `cold-display-guard-runtime` and `cold-display-guard-api` recreated and started |
+| 2026-05-29 | Phase 4 | Final Code Review Agent | Reviewed full v1.2 implementation | Verdict fail: extra trash fallback was suppressed, tracker could seed from outside-source motion, and docs named a non-existent backend |
+| 2026-05-29 | Phase 4 | Main Agent | Fixed final review findings | Added regression tests, adjusted trash fallback budgeting, rejected outside-before-source trajectories, and renamed docs to `TrajectoryTracker` |
+| 2026-05-29 | Phase 4 | Main Agent | Re-ran local full verification | Python 70 tests passed; frontend 22 tests passed; Vite build passed; no heavy model dependency matches |
+| 2026-05-29 | Phase 4 | Main Agent | Re-synced and redeployed fix to `xiaozheng@192.168.5.206` | Runtime/API rebuilt and restarted from the fixed code |
+| 2026-05-29 | Phase 4 | Main Agent | Verified remote runtime after redeploy | Containers are up, API health returns `status=ok`, diagnostics contain `disposal_evidence` and `diagnostics.trajectory` |
 
 ### Test Results
 
@@ -224,6 +234,13 @@
 | 2026-05-29 | `PYTHONPATH=src python3 -m unittest tests.test_main tests.test_vision -v` | pass | 26 runtime/vision tests passed after phase 3 |
 | 2026-05-29 | `PYTHONPATH=src python3 -m unittest discover -s tests -v` | pass | 68 full Python tests passed after phase 3 |
 | 2026-05-29 | `rg -n "ultralytics|torch|onnxruntime|openvino|opencv|cv2|numpy" src tests pyproject.toml` | pass | No matches; command exited 1 because no heavy vision/model dependency was found |
+| 2026-05-29 | `PYTHONPATH=src python3 -m unittest discover -s tests -v` | pass | 70 full Python tests passed after final review fixes |
+| 2026-05-29 | `node --test web/test/zone-state.test.js` | pass | 22 frontend model tests passed |
+| 2026-05-29 | `cd web && pnpm build` | pass | Vite production build passed |
+| 2026-05-29 | `rg -n "ultralytics|torch|onnxruntime|openvino|opencv|cv2|numpy" src tests pyproject.toml` | pass | No matches; command exited 1 because no heavy vision/model dependency was found |
+| 2026-05-29 | `docker compose ... build cold-display-guard-runtime cold-display-guard-api && docker compose ... up -d --no-deps ...` on remote | pass | Runtime/API rebuilt and restarted after final fixes |
+| 2026-05-29 | `curl --max-time 5 http://192.168.5.206:19080/api/manage/health` | pass | `status=ok`, `runtime_status=running` |
+| 2026-05-29 | Remote diagnostics schema check script | pass | `has_disposal_evidence=True`, `has_trajectory=True` |
 
 ### Bug Loop
 
@@ -238,6 +255,44 @@
 | Phase 3 | Capture-error diagnostics rows lack `disposal_evidence` and `diagnostics.trajectory` schema | Added regression test and wrote empty evidence plus trajectory error diagnostics on capture failure | Resolved; testing agent and local full Python suite passed |
 | Phase 3 | Runtime tests patch `TrajectoryTracker` with `create=True`, which can mask missing imports | Removed `create=True` and asserted fake tracker observe calls | Resolved; testing agent and local full Python suite passed |
 | Phase 3 | `disposal_evidence_payloads()` accepts `list[object]` but blindly calls `asdict()` | Narrowed helper signature to `list[DisposalEvidence]` | Resolved; testing agent and local full Python suite passed |
+| Phase 4 | Remote config patch inline heredoc lost the empty-string value for `yolo_model_path` | Switched to an scp-uploaded Python patch script instead of repeating inline quoting | Resolved; remote config patch reported 13 runtime keys patched |
+| Phase 4 | Source-specific evidence disabled all generic fallback trash deposits | Subtract only the count of evidence-discard events from `remaining_trash_deposits`; add regression test for evidence plus `trash_deposit_count=2` | Resolved; targeted regression and full Python suite passed |
+| Phase 4 | Tracker could keep an outside-source blob as the first point before source motion | Track outside-before-source contamination, reject later source noise with `motion_started_outside_source`, and never append those outside points | Resolved; targeted regression and full Python suite passed |
+| Phase 4 | `docs/project.md` referenced non-existent `MotionTrajectoryBackend` | Changed architecture docs to name the implemented `TrajectoryTracker` | Resolved; `rg` no longer finds the stale name in project docs/README/src/tests |
+
+## 2026-05-29 Phase Completed: Phase 4 - Documentation, Verification, And Deployment
+
+Status: complete
+
+Files Changed:
+- `.gitignore`
+- `README_zh.md`
+- `docs/project.md`
+- `findings.md`
+- `memories.md`
+- `progress.md`
+- `task_plan.md`
+- `src/cold_display_guard/engine.py`
+- `src/cold_display_guard/vision.py`
+- `tests/test_engine.py`
+- `tests/test_vision.py`
+
+Tests:
+- `PYTHONPATH=src python3 -m unittest discover -s tests -v`: pass, 70 tests
+- `node --test web/test/zone-state.test.js`: pass, 22 tests
+- `cd web && pnpm build`: pass
+- `rg -n "ultralytics|torch|onnxruntime|openvino|opencv|cv2|numpy" src tests pyproject.toml`: pass, no matches
+- Remote Docker rebuild/restart: pass
+- Remote API health: pass
+- Remote diagnostics schema check: pass
+
+Notes:
+- Final review findings were fixed before redeploy.
+- Remote sync continued to exclude `config/example.toml` to preserve camera/calibration settings.
+- Remote diagnostics check intentionally printed only schema booleans and safe trajectory keys.
+
+Risks:
+- v1.2 is still heuristic motion tracking. Live precision should be tuned from diagnostics, and future YOLO integration should continue using the same `disposal_evidence` contract.
 
 ## 2026-05-29 Phase Completed: Phase 3 - Runtime Integration
 

src/cold_display_guard/engine.py

@@ -25,18 +25,15 @@ class BatchEngine:
         previous_zone_counts = dict(self._zone_counts)
         remaining_trash_deposits = observation.trash_deposit_count
         used_disposal_evidence: set[int] = set()
-        has_source_specific_disposal = self._has_confirming_disposal_evidence(observation.disposal_evidence)
-        if has_source_specific_disposal:
-            remaining_trash_deposits = 0
 
         events.extend(self._expire_pending_disposal(observation.ts))
-        events.extend(
-            self._apply_disposal_evidence(
-                observation.ts,
-                observation.disposal_evidence,
-                used_disposal_evidence,
-            )
+        evidence_events = self._apply_disposal_evidence(
+            observation.ts,
+            observation.disposal_evidence,
+            used_disposal_evidence,
         )
+        remaining_trash_deposits = max(0, remaining_trash_deposits - len(evidence_events))
+        events.extend(evidence_events)
         trash_events = self._apply_trash_deposits(observation.ts, remaining_trash_deposits)
         remaining_trash_deposits = max(0, remaining_trash_deposits - len(trash_events))
         events.extend(trash_events)
@@ -82,13 +79,13 @@ class BatchEngine:
             self._zone_counts[zone_id] = new_count
 
         newly_pending_count = max(0, len(self.pending_disposal) - pending_count_before_zone_transitions)
-        events.extend(
-            self._apply_disposal_evidence(
-                observation.ts,
-                observation.disposal_evidence,
-                used_disposal_evidence,
-            )
+        evidence_events = self._apply_disposal_evidence(
+            observation.ts,
+            observation.disposal_evidence,
+            used_disposal_evidence,
         )
+        remaining_trash_deposits = max(0, remaining_trash_deposits - len(evidence_events))
+        events.extend(evidence_events)
         trash_deposits_to_apply = remaining_trash_deposits
         if remaining_trash_deposits > 0 and newly_pending_count > 1:
             trash_deposits_to_apply = max(remaining_trash_deposits, newly_pending_count)
@@ -283,9 +280,6 @@ class BatchEngine:
             events.append(self._event("batch_discarded", when, batch, severity="info"))
         return events
 
-    def _has_confirming_disposal_evidence(self, disposal_evidence: list[DisposalEvidence]) -> bool:
-        return any(self._is_confirming_disposal_evidence(evidence) for evidence in disposal_evidence)
-
     def _is_confirming_disposal_evidence(self, evidence: DisposalEvidence) -> bool:
         if evidence.confidence < DISPOSAL_EVIDENCE_CONFIDENCE_THRESHOLD:
             return False

src/cold_display_guard/vision.py

@@ -83,6 +83,7 @@ class _TrajectoryCandidate:
     last_sample_at: datetime | None = None
     points: list[_MotionPoint] | None = None
     source_motion_seen: bool = False
+    pre_source_motion_seen: bool = False
 
     def __post_init__(self) -> None:
         if self.points is None:
@@ -456,13 +457,22 @@ class TrajectoryTracker:
                 available_source_blobs = [blob for blob in source_blobs if blob.blob_id not in consumed_blob_ids]
                 if not available_source_blobs:
                     return "ambiguous_motion_track"
+                if candidate.pre_source_motion_seen:
+                    return "motion_started_outside_source"
                 candidate.source_motion_seen = True
                 point = _nearest_point(region_center(candidate.source_region), available_source_blobs)
             else:
                 available_blobs = [blob for blob in blobs if blob.blob_id not in consumed_blob_ids]
                 if not available_blobs:
                     return None
-                point = _nearest_point(region_center(candidate.source_region), available_blobs)
+                candidate.pre_source_motion_seen = True
+                if any(
+                    region_contains(self.trash_region, blob.x, blob.y, margin=self.settings.trajectory_trash_entry_margin)
+                    for blob in available_blobs
+                    if self.trash_region is not None
+                ):
+                    return "motion_started_outside_source"
+                return None
         else:
             previous = candidate.points[-1] if candidate.points else None
             available_blobs = [blob for blob in blobs if blob.blob_id not in consumed_blob_ids]

task_plan.md

@@ -85,12 +85,12 @@ Create and evolve an independent git project under `~/Code` for monitoring food
 
 ### Stop Conditions
 
-- [ ] v1.2 所有阶段完成。
-- [ ] 必要 Python 测试通过。
-- [ ] 前端测试或构建在受影响时通过。
-- [ ] `docs/project.md` 记录 v1.2 架构、配置、运行方式和关键决策。
-- [ ] 没有 blocking bug 或未处理的高风险问题。
-- [ ] 如果同一问题连续 3 次修复仍失败，暂停并报告原因、已尝试方案和建议下一步。
+- [x] v1.2 所有阶段完成。
+- [x] 必要 Python 测试通过。
+- [x] 前端测试或构建在受影响时通过。
+- [x] `docs/project.md` 记录 v1.2 架构、配置、运行方式和关键决策。
+- [x] 没有 blocking bug 或未处理的高风险问题。
+- [x] 如果同一问题连续 3 次修复仍失败，暂停并报告原因、已尝试方案和建议下一步。
 
 ### Phases
 
@@ -99,7 +99,7 @@ Create and evolve an independent git project under `~/Code` for monitoring food
 | 1 | complete | 建立 `disposal_evidence` 数据契约并让状态机优先按来源区域丢弃 | `Observation` 支持 evidence；engine 能按 `source_zone_id` 精确关闭 pending batch；同帧移除+evidence 有回归测试；旧 `trash_deposit_count` 仍可兜底 |
 | 2 | complete | 实现无 YOLO 依赖的轻量轨迹检测 | synthetic frame 测试覆盖源区域到垃圾桶、非源区域运动、未到垃圾桶、单帧反光、多候选互不串扰；不引入模型依赖 |
 | 3 | complete | 集成 runtime 配置、诊断和候选窗口加速采样 | `main.py` 写入 `disposal_evidence` 与 trajectory diagnostics；配置默认 `trajectory_enabled=true`、`yolo_enabled=false`；候选活跃时使用更短采样间隔 |
-| 4 | pending | 文档、全量验证和部署准备 | README/project/progress 更新；Python 全量测试通过；前端测试/构建按影响范围验证；远端部署命令和风险记录清楚 |
+| 4 | complete | 文档、全量验证和部署准备 | README/project/progress 更新；Python 全量测试通过；前端测试/构建按影响范围验证；远端部署命令和风险记录清楚 |
 
 ### v1.2 Decisions
 

tests/test_engine.py

@@ -250,6 +250,33 @@ class BatchEngineTests(unittest.TestCase):
         self.assertEqual([(event["event"], event["zone_id"]) for event in events], [("batch_discarded", "4")])
         self.assertEqual([batch.zone_id for batch in engine.pending_disposal], ["1"])
 
+    def test_extra_trash_deposits_still_fallback_after_matching_disposal_evidence(self) -> None:
+        settings = EngineSettings(
+            camera_id="test_cam",
+            max_dwell_seconds=1200,
+            trash_confirmation_seconds=120,
+            zone_ids=("1", "2"),
+        )
+        engine = BatchEngine(settings)
+        engine.process(obs(self.t0, {"1": 1, "2": 1}))
+        engine.process(obs(self.t0 + timedelta(seconds=1200), {"1": 1, "2": 1}))
+        engine.process(obs(self.t0 + timedelta(seconds=1300), {"1": 0, "2": 0}))
+
+        events = engine.process(
+            obs(
+                self.t0 + timedelta(seconds=1310),
+                {"1": 0, "2": 0},
+                trash=2,
+                disposal_evidence=[disposal_evidence("1")],
+            )
+        )
+
+        self.assertEqual(
+            [(event["event"], event["zone_id"]) for event in events],
+            [("batch_discarded", "1"), ("batch_discarded", "2")],
+        )
+        self.assertEqual(engine.pending_disposal, [])
+
     def test_same_observation_removal_and_disposal_evidence_discards_newly_pending_batch(self) -> None:
         settings = EngineSettings(
             camera_id="test_cam",

tests/test_vision.py

@@ -474,7 +474,39 @@ class VisionTests(unittest.TestCase):
             rejected.extend(diagnostics["rejected"])
 
         self.assertEqual(all_evidence, [])
-        self.assertTrue(any(item["reason"] == "missing_source_motion" for item in rejected))
+        self.assertTrue(any(item["reason"] == "motion_started_outside_source" for item in rejected))
+
+    def test_motion_before_source_motion_cannot_seed_later_trash_evidence(self) -> None:
+        source = Region("source", ((0.05, 0.35), (0.25, 0.35), (0.25, 0.65), (0.05, 0.65)))
+        trash = Region("trash", ((0.72, 0.35), (0.95, 0.35), (0.95, 0.65), (0.72, 0.65)))
+        tracker = TrajectoryTracker(
+            [source],
+            trash,
+            RuntimeVisionSettings(trajectory_sample_interval_seconds=0.0, trajectory_min_points=3),
+        )
+        now = datetime(2026, 4, 28, 10, 0, tzinfo=timezone.utc)
+
+        tracker.observe(solid_frame(80, 80, 40), now, {"source": 1})
+        all_evidence = []
+        rejected = []
+        frames = [
+            frame_with_motion_patch(80, 80, (34, 36)),
+            frame_with_motion_patch(80, 80, (16, 36)),
+            frame_with_motion_patch(80, 80, (34, 36)),
+            frame_with_motion_patch(80, 80, (50, 36)),
+            frame_with_motion_patch(80, 80, (66, 36)),
+        ]
+        for index, frame in enumerate(frames):
+            evidence, diagnostics = tracker.observe(
+                frame,
+                now + timedelta(seconds=index + 1),
+                {"source": 0},
+            )
+            all_evidence.extend(evidence)
+            rejected.extend(diagnostics["rejected"])
+
+        self.assertEqual(all_evidence, [])
+        self.assertTrue(any(item["reason"] == "motion_started_outside_source" for item in rejected))
 
     def test_trajectory_diagnostics_include_per_candidate_events(self) -> None:
         source = Region("source", ((0.05, 0.35), (0.25, 0.35), (0.25, 0.65), (0.05, 0.65)))