@trillboards/edge-sdk 0.2.104 → 0.2.107

package/deploy/detector/README.md

@@ -133,6 +133,39 @@ states the box is in. The service does not exit when nothing was injected: all
 three supervisors restart on a non-zero exit, and this file's own rule for a
 required input that has not arrived is *start, and report the reason*.
 
+#### Who actually supplies it — added 2026-08-26, because until then NOBODY did
+
+This section described a parameter that existed on both spellings and was passed
+on neither. `TRILLBOARDS_DETECTOR_RAM_BUDGET_MB` appeared in this file, in
+`cli.py`, in `modelpath.py` and in the tests, and in **zero** of the six
+supervisor definitions — so on every fielded box the "not injected" row of the
+table above was not one of two states, it was the only one, and every delivered
+specialist was permanently `idle`.
+
+`ram-budget.js`, in this directory, is the supplier. It is run by the three
+installers, not by this service — the ceiling must stay something this process is
+TOLD — and it derives the number from the box rather than picking one, because
+`docs/runbooks/sensing-model-library.md` §2 is arithmetic about a SET on a
+specific machine and the two measured field boxes are 9.8 GB and 31.4 GB. It
+subtracts the on-device VLM's measured working set, the connector process and the
+capture lane from physical memory, floors the result at the pinned model's own
+footprint (below that `ModelRegistry.__init__` raises and the supervisor
+crash-loops, which is strictly worse than being constrained), and prints the whole
+subtraction with `--explain` so an installer log says why.
+
+Three transports, one number, because the three supervisors differ in what they
+can carry:
+
+| platform | how | why not the other one |
+| --- | --- | --- |
+| Linux | `TRILLBOARDS_DETECTOR_RAM_BUDGET_MB` in `/etc/trillboards/env` | the unit is a static file `install.sh` copies verbatim; there is nothing to template, and both units already declare `EnvironmentFile=-` |
+| macOS | `--ram-budget-mb` in the plist's `ProgramArguments` | the plist IS generated by `setup-detector.sh`, from a template |
+| Windows | `--ram-budget-mb` in the Scheduled Task's argument string | a Scheduled Task action has no environment block, and a User-scope variable would be global machine state |
+
+On every platform, **failing to derive injects nothing** rather than a fallback —
+the "not injected" row stays reachable, and it is the honest thing to report about
+a box we could not measure.
+
 **Eviction is a last-resort safety valve, not the allocator.** It logs a WARNING
 naming the box as oversubscribed against what it was allocated, and `/health`
 carries a top-level `evictions` count — because a service quietly thrashing

package/deploy/detector/bench/litert-twin-parity.json

@@ -20,7 +20,18 @@
     "its scores agree to 7.2e-08, but row-wise the two disagree and lexsorting by box does",
     "not collapse it. The model returns ZERO detections above 0.25 on 40 real production",
     "frames, so there is no above-threshold sample to compare and row-level parity is",
-    "UNVERIFIED. It is not published. Settle it on a corpus the model actually fires on."
+    "UNVERIFIED. It is not published. Settle it on a corpus the model actually fires on.",
+    "",
+    "2026-08-26: the 22nd twin, gh_fall_tonlong_v2, is the first whose LiteRT input is",
+    "NHWC [1,640,640,3]. The line above saying an ultralytics detect export is NOT NHWC",
+    "described the TOOLCHAIN the first 21 were exported with, not LiteRT: onnx2tf",
+    "transposes to channels-last by design and ultralytics 8.3.145 leaves it that way.",
+    "verify_litert_twin_parity.py used to ASSERT NCHW and refused a correct twin; it now",
+    "reads the input layout off the graph -- the same rule it already applied to outputs",
+    "-- and feeds each runtime the SAME IMAGE in the layout that runtime declares. The",
+    "warning that motivated the assertion still stands: transposing because 'LiteRT is",
+    "channels-last' feeds a rotated image to a model that wanted NCHW, and every residual",
+    "then reads as export noise."
   ],
   "measured_at": "2026-08-25",
   "tolerances": {
@@ -308,6 +319,21 @@
       "reason": "ok",
       "onnx": "weapon_gunknife_11n.onnx",
       "litert": "weapon_gunknife_11n.tflite"
+    },
+    {
+      "layout": "anchor_4_plus_nc",
+      "imgsz": 640,
+      "box_scale_applied": 640.0,
+      "box_residual_px": 0.0007171630859375,
+      "box_residual_unscaled": 646.9068156480789,
+      "cls_residual": 1.3665612286217765e-07,
+      "pass": true,
+      "reason": "ok",
+      "onnx": "gh_fall_tonlong_v2.onnx",
+      "litert": "gh_fall_tonlong_v2.tflite",
+      "litert_input_layout": "nhwc",
+      "measured_at": "2026-08-26",
+      "note": "The 22nd twin, added 2026-08-26, and the FIRST whose LiteRT input is NHWC. gh_fall_tonlong had no parity measurement at all until now: upstream ships `Violence` at slot 0 and `violence` at slot 6, normalize_class_name folds them to one name and ModelVocabulary refuses at construction, so neither format could be probed. Both were re-exported from ok.pt (sha256 1e12d89198e28d7c650ba86886c511fecd9f70b9d7f8bf985e76704690e58dd9) with the slots disambiguated by INDEX -- violence_slot0 / violence_slot6 -- which preserves upstream truth and asserts no semantics we have not measured."
     }
   ]
 }