chore(dev): release 0.2.0

[github-actions]

🤖 I have created a release beep boop

0.2.0 (2026-08-19)

⚠ BREAKING CHANGES

• remove the quadruped line

Features

• --stabilise on hydranet-scene, and the Pillow constant the stubs dropped (f84429c)
• --vocab retail on hydranet-scene, so a display case stops being an oven (a4bc777)
• product is a detection question, so take it out of the dense head (e92b079)
• a column supplement arm, and the held-out cameras that judge it (8184010)
• a control for the FCOS towers at half depth (9a23366)
• a crop encoder whose attribute training bought association it was not asked for (e79421d)
• a depth head, the third head family the registry was built for (0f8f8ea)
• a loader for rendered depth, and a 300x faster semantic snap (e1f2bdf)
• a night person veto that removes a box whose pixels never moved (359f2dc)
• a python TensorRT sweep, and bench deps that survive a resync (5b6d380)
• a second retail taxonomy, for naming objects rather than free space (31079fa)
• a shape for the classes whose name is already the claim (e1ca530)
• a stabiliser that has stopped working says so instead of coasting (dfa42e5)
• a vanishing point says how well its lines actually meet (0a6f276)
• accept a directory of stills, for the cameras nobody kept the video of (c95291f)
• adopt the static-plate builder the shared tree already relies on (56acd77)
• adopt the static-share scorer for person pseudo-boxes (f3525aa)
• alerts become training data the day they are raised, not later (46793c9)
• arrows steer in movement mode instead of doing nothing (dc419ec)
• batch03 annotation, class weights for the class that was eating its neighbours (177e9fc)
• body height while walking, and a forwarder that survives the cable coming out (6a6595b)
• calibrate a fixed camera from its floor tiles, distortion first (cb39a27)
• check fixed_weights against the heads that exist (5d9e448)
• check the site split against its own rules, from the record (8b7e221)
• column is sourceable from site after all, at a threshold it had to earn (f2ee5bd)
• deploy/robot, the quadruped surface of a repo that was only retail (87025fd)
• derive free space from terrain when a config drops the traversability head (573290c)
• draw the floor panel as a room rather than as a plot (362a224)
• draw the floor scene like something you would look at, not a plot (ce8445f)
• draw the person mesh in the panel people actually look at (4cbc6f3)
• E-prep run once, and the sonar anchor it disqualifies (accd2c8)
• evaluate the crop encoder on what mean accuracy hides (0eeacd5)
• fold the segmentation argmax into the exported graph (bb6f306)
• give the text-embedding head the matrix it was always documented to have (d9bb6b1)
• in-domain person supervision from the Grounding DINO site batch (3e37368)
• join three night measurements into a per-camera person verdict (caae812)
• label a fixed camera by consensus, for when nobody can correct a mask (f0d9e0b)
• label rendered poses from the geometry, not from where the camera was put (7642de1)
• let the ROS live view choose the interface it serves on (af9e7c1)
• let the segmentation loss weight its classes (4c29036)
• make the detection vocabulary data instead of weights (126298a)
• make torch.compile an opt-in training knob (cf17644)
• make track ground truth a judgement instead of a labour (a3a0fa2)
• measure identity before building anything that claims to fix it (078f0e9)
• measure the model on the store, without labels (faa6198)
• measure the val set's own error bar before buying more of it (5a39b52)
• measure what int8 actually costs, and stop asserting it (fdfee2e)
• mirror runs/ into a local MLflow index; files stay the truth (888d4bb)
• move cameras between splits in every batch at once, or not at all (2f8fe95)
• name the classes one dataset has and a bigger one does not (bc43b57)
• narrow the detection classes at export, not at training (4fb05ff)
• night person filter CLI, and the acceptance criteria as tests (95e9be1)
• offline non-causal tracking turns a clip into a handful of merge decisions (4e5ff7f)
• one answer per shopper, with the number that says how much to trust it (3aaeefe)
• one camera and one light rig for anything that draws a mesh (0a78138)
• one detection head can hold person and boxed_stock, if the loss stops lying (12b236d)
• one plate calibrates orientation; the metre needs a known length (e003b54)
• optional CUDA graph capture in the live viewer (b60f79e)
• people-flow analytics, as post-processing rather than a fourth head (d594db7)
• pose pilot, the first current through the tier-2 keypoint slot (4538867)
• pre-label with SAM 3 the classes no public dataset supplies (74dc2d0)
• procedural meshes for the scene, sized from what was measured (f0558ef)
• push the retail canvas to 640x1120, the lever that measured largest (ee1bdc1)
• put a wall clock on every security event (5dac1fb)
• quad layout renders the depth head on a fixed metre ramp (84c90e5)
• RAP v2 lands on the attribute pipeline, and cameras onboard from one plate (935ef42)
• re-sweep column per camera, and keep the identities this time (2388b88)
• read merchandise out of SAM 3 as instances, not as pixels (7cd071e)
• rebalance the site run, because the class was suppressed rather than unlearned (404a7e4)
• refuse a dirty tree at the start of training, not at release (3375696)
• remove the quadruped line (eca3814)
• render HM3D at a measured camera pose, without habitat-sim (b56ed0a)
• report what each per-class IoU stands on, not just the IoU (e4655a0)
• say at config time which classes no dataset can produce (a716c37)
• say what best.pt gave up on the heads it was not selected for (64a7cbf)
• security events are geometry and time over tracks, and none of them is a head (10cb1bb)
• serve_pilot -- 16 streams end to end hit 16 x 14.6 fps, post is the frontier (cbe1aef)
• serving package -- the H2D recovery and the per-camera serving state (0086a87)
• shelf coverage at native resolution, and a daylight threshold that was guessed (70759db)
• ship the fleet aggregator, because nanmean would hide the busiest camera (66b1ad4)
• stop training when it stops improving, and score mAP less often (dcc4b2f)
• teleop control interface, the whole command set the spec documents (adb579a)
• the 96-stream 15 fps serving target, and the sweep that will measure it (cc80fc3)
• the configs behind today's four runs, with their empty channels pinned (3a8608c)
• the flicker ruler, measured before any smoothing exists to help (e94781e)
• the HM3D-rendered depth experiment config for the CCTV line (933c93c)
• the official 0x0901 telemetry the beta comm spec documents, wired in (b40c6fb)
• the robot's model plus a depth head, one change from the deployed config (1e5454f)
• the teacher's error bar, measured on NYUv2 instead of waiting for the robot (28e63e8)
• the whole stack as a video, and site person boxes whose second gate failed its own test (45b6732)
• three paired seeds for the retail+security run, and the config check that saves them (298b727)
• three stabilisation switches, measured against the flicker ruler (d8542fd)
• train on the shop, not on photographs of other buildings (ae554c0)
• zone proposals from the fleet calibration, for a human to confirm (d94faf5)

Fixes

• product needs its own score threshold, because goods are not furniture (b0ab253)
• a config gained class names and its test did not (cf99f18)
• a det_vocab checkpoint names its own boxes; --vocab is for COCO heads (554e5af)
• a detector that returned nothing has an mAP, and it is zero (517b943)
• a palette for retail_surfaces, which crashed three seed runs on epoch 1 (1755f0d)
• a report says what decided its numbers, not where anything lives (73e96ed)
• a review sheet with no crops is not a review sheet (a01d1d6)
• a run records the directory it wrote to, not the one it was asked for (84c9f82)
• a second run in one process wrote its log into the first run's directory (93f57cc)
• a session directory should mean frames were written (26def9e)
• adopt 3-of-4 tranches as the column supply rule, and apply it (7ae9ee9)
• an all-ignore batch has no mean IoU, and should say so without warning (592a0a6)
• an instance box the size of a counter is not an instance (b61ca0c)
• an output directory named after a filename is not an identity (ed0b65d)
• bound the ffprobe call, so one unreadable clip cannot stall a batch (537c341)
• check augmentation settings where they are read, and ratchet the baseline down (559ffa3)
• claim a run directory when it is resolved, not when meta.json lands (d5858cf)
• ffprobe refuses the section name every video path asks for (550138a)
• four annotations that were wrong, not merely unproven (666a396)
• guard the by-path loader against None specs, paying the scripts/ ratchet back to 18 (783fc5b)
• ignore the .plan engines the pro6000 sweep leaves behind (e201999)
• let a point vote only for the line direction its gradient supports (08763f4)
• make assets/ an allowlist, so a store frame cannot be added by accident (20ebaf0)
• neither Jetson script started on a CUDA 13 board (547b14c)
• no module unpickles a checkpoint outside load_checkpoint (4e54bb6)
• one clip loop, and it refuses to guess whether to correct the lens (eb7adb6)
• pay the src/ type debt back past where it started, 11 -> 8 (e468209)
• pin the type ratchet's interpreter, because the number moved without the code (a7d8410)
• probe the frame rate that was recorded, not the one the container claims (cef93b2)
• raise when ffmpeg stops early instead of calling it the end of the video (40a553c)
• reach_to_shelf reads what the hand is over, not the pixel the hand hides (9e6818c)
• reading a finished run must not require the validator written after it (ea75264)
• Recorder is a context manager, which five tests had been proving it needed to be (854fe0a)
• refuse a resplit that would leak into a train-only supplement (23f88ee)
• refuse a run that trains on the cameras it will be scored on (8bd195f)
• scope a numpy-2.5 stub false positive so the scripts ratchet reads true (5e2d6e1)
• score each validation set separately, because pooling hid a working model (1f88a76)
• select this run on the store's val frames, not on ADE20K's hallucinations (fe08559)
• shut the video encoder down on every path, and check that it succeeded (efd4345)
• spell the PIL constants the way a static checker can see them (a3778b0)
• stop caveating void, which nobody trains (6d3cc45)
• the acceptance test passed a camera pointed at the ceiling (af2922c)
• the annotation gate leaked a file handle per mask it refused (93351db)
• the dashboard page was cacheable, so a redeploy could be invisible (7373b45)
• the minority-source check confused expressible with present (75cd43e)
• the Orin is the retail line's board, not the quadruped's (e77a219)
• the panel was naming boxes from a table the head does not use (47c47e3)
• the tracker matched against a prediction, and for a standing shopper that is jitter (759bb41)
• the type ratchet was rewriting the venv it measured (b4e8c81)
• three tables that named things they do not describe (5809d4f)
• two of three teleop axes were inside the robot's deadzone, plus arrow-key posture (febd464)
• write checkpoints through a rename, so a preemption cannot destroy the last one (4f01269)

Performance

• batch the EMA update instead of launching a kernel per tensor (590e35f)
• raise workers and channels_last from a measured ceiling, not a guess (9e4378e)
• SAM 3 was re-encoding the same image 44 times a frame (b030275)
• snap semantic colours over distinct values, not pixels (163289d)

Refactoring

• 255 has one definition, and labels.py finally reaches the places it named (458fdea)
• check one dataset entry in one function (1b4aa8b)
• deploy/ is where the model runs, so annotation is not a deploy target (cb07537)
• drop the flat BEV panel, superseded as the default since 362a224 (db1d0e0)
• drop the last private source_fps, now that probe reads the same field (a0d9471)
• frame selection is library code, not one script importing another (35dc2f8)
• give prepare-cocostuff a build_parser, and its first tests (25f8f1c)
• give the preprocessing constants a home below everything (66cb570)
• give the two scene payloads a shape a checker can hold you to (846d920)
• group the quadruped line into its own scripts/ and configs/ subdirectories (bd1ecaa)
• let the preview be a scene, not a second renderer (8370d5c)
• lift five concerns out of the Trainer constructor (89afb98)
• make the network iterate its heads instead of unrolling detection (9a2921d)
• move per_image_confusions into the package as engine/confusion (80ce609)
• move EMA and the optimiser schedule out of the trainer module (025a34a)
• move the ByteTrack forward pass into the package as analytics/bytetrack (c6c67a0)
• move the live view's own work into the package (c407261)
• narrow the Optionals where the invariant actually holds (0c2f39f)
• one resolve for both SAM 3 prompt tables (0187ca5)
• one IGNORE, and a check that the configs still mean it (ed92829)
• one IoU for the tracker and for the metric that scores it (be44029)
• read the detection head through one accessor that checks it (9ee49a1)
• seed the BiFPN passes from their input, not from None (2d8e336)
• split analytics/events into a package, on the lines it already drew (b09549c)
• split Trainer.__init__ into the seven phases it already had (868775d)
• split check_split by the four questions it asks (63f1b4a)
• split evaluate into accumulate and reduce, and fix two annotations (abac49f)
• split prepare-ade20k's main into clear, link and report (8755ad6)
• split the 3D panel renderer along the seams it already had (ed4eb51)
• split the stage contract by what a consumer reads, and put it in force (25c0976)
• the plate-calibration pipeline moves into the package, dissolving two script pairs (6a2f634)
• the third copy of the frame decoder goes through data/video (2f62791)
• thin the scene and infer-video mains, and test what came out (4b60381)
• video decoding is a data question, so it stops living in the CLI (7466604)

Documentation

• coco10 is share 1.0, and reading it as 0.10 invents a contradiction (eb49467)
• experiment in meta.json is not an identifier, and the sweep proves it (d891f01)
• masks means class ids, and ToTensor's int64 cast is that contract (24da737)
• 2.27x compute margin on the 96-stream target; the frontier is the copy path (1c7b6bc)
• a background plate recovers a camera's orientation and cannot set its metre (06fdf72)
• a held-out camera for the README figure, and the caveat it removes (4e16ae9)
• a run whose length a metric decides is not comparable to one it does not (ae30c52)
• an audit found the live docs were arguing from numbers three retractions old (b74f264)
• both labelled clips done -- coexistence evidence, not crop resemblance (5bf57a0)
• clear the artefacts the 2026-08-19 seventeen-to-nine merge left behind (69de1d9)
• correct a claim about checkpoint selection, which was backwards (ef2b54d)
• correct the night verdict to 13 counter-examples and one real person (7c01da9)
• current outputs at a glance, dated, inside the contract section (ad6dfa8)
• cutting v1 found the baseline everyone had been quoting was three epochs (8d80585)
• delete the 2026-08-17 session board, which asked to be deleted (ec316e9)
• device splits into laptop/tablet/phone where the prompt evidence pointed (b0d2047)
• drop two figures RETAIL_SCOPE.md was already referencing into a hole (268e9f8)
• EMA wins unconditionally, tracks replace raw boxes, composite waits for serving (5d821bd)
• fit the lens before the pose, and prefer tiles to people (d08db25)
• fleet calibration -- 83 percent automatic, failures are mount types (0b1665f)
• geometry/ now has a second copy in production, on a runtime it predates (b8700c8)
• handoff closes with v4.2 at the user's gate (541fc3a)
• handoff for the retail+security line, written before the context ran out (1a836ad)
• journal the column leak, the retraction, and two negative results (635e3f1)
• label the quadruped line as secondary where it lives, without deprecating it (19d9bbb)
• lead with the deployment that exists, not the one the repo was named for (38f8b56)
• name the second stage, and the order the retail work has to be built in (3f50700)
• nine, not eight -- and the ninth is a record that agrees with itself and is wrong (53b2250)
• one deployment target, and say which of the three configs is the product (93b5f06)
• one of the robot copy's five patches deletes a guard rather than re-spelling one (9dcc759)
• planner handoff -- the live state a successor cannot reconstruct (5f49baa)
• priors are a tape measure to within 5-8 percent (3a34ca2)
• put the two ratchet warnings where someone editing the file will hit them (46e0f63)
• rank annotation by what no teacher can supply, not by what the LiDAR cannot see (d867b67)
• RAP v2 trains tonight; AgeOver60 has no source, staff finally does (5972c1c)
• reach_to_shelf now fires on touch, and it found the missed shopper (2e3719a)
• record a mechanism that is true and a proxy for it that is not (1fd51db)
• record that R3 is unsatisfied, rather than leaving it looking met (257eca1)
• repoint the three figures that f9d4fcf left dangling (bb16978)
• research-phase licence ruling, the day's arrivals, and the VLM pilot verdict (dced4ba)
• retire four claims the code has since disproved (c30a360)
• retrain the training guide on the line this project actually ships (2df7fe8)
• rule 4 says detection for a reason, found by generalising it anyway (df85b53)
• say what the split hash is for, in the argument that says it (1a5b56f)
• scene.py pointed at a script that is not in the tree (f62934e)
• separate the two detection-narrowing knobs, since one is a trap (5a27819)
• seventeen documents where the boundaries had stopped matching the work (8b37959)
• the attribute-only retrain hurt the embedding it exists to serve (9840dd4)
• the bootstrap does not decline to label a podium, it labels it wrong (c38f414)
• the citation file was the last thing still calling this a quadruped project (e80e2a1)
• the column supplement leaked, and re-sweeping for more cameras found none (959dffd)
• the disposition log exists, and what it must never be read as (72cd8c1)
• the export contract changed, so the places that state the old one had to (bf69d2a)
• the fleet census by eye -- five sideways mounts and a missed panorama (5e97b37)
• the fourth video layout, and the experiment that questions this file's own conclusion (307d03d)
• the GB10 figures are +-15%, and the table did not say so (ec9ce93)
• the immovable classes are decided per object, not per pixel (8917f15)
• the keypoints contract cited a function that was never written (2c2188d)
• the measured frame, and why trtexec alone would revert --argmax-seg (2358cce)
• the missing retail_surfaces scheme is the design, not a gap (0da8588)
• the module that justified keeping 80 detection classes was never written (f69d258)
• the night gap belongs to one camera -- the swap moved the failure (3432f38)
• the nine was right by accident, and the worse case is the one it missed (673ef3e)
• the occupancy research direction, and a handoff for the robot line (5853d08)
• the package still told PyPI it was for quadruped robots (7f5fea9)
• the panel drawn from a site clip, with the scores it was drawn at (3575778)
• the person teacher swap, the near-labelled clip, and what a plate calibrates (d9224c8)
• the pose slot powers on, and flicker gets its ruler (6d9f7c1)
• the pre-processing contract said the opposite of what the graph does (d9c1a45)
• the README's lead image showed a panel the code no longer draws (19daa58)
• the record fix covers the failure that announces itself, not the one that does not (fb86b3d)
• the retrain's first seed cuts static flicker by a third (c4dfc8a)
• the robot the annotation priority was ranked for has no LiDAR (025d100)
• the session board for 2026-08-17, while three sessions still share the tree (aa192af)
• the site test split, and why a 0.77 mIoU and a 0.00% class are both true (26149f0)
• the two missing classes are no longer missing, and one prediction was wrong (f327e96)
• thirteen counter-examples, and static share only points one way (e8ac621)
• three seeds agree -- in-domain person lands at no cost to any other head (a4e7390)
• transfers leave the critical path; python post inherits the bottleneck (655329b)
• translate onboard_camera.py, the last Chinese-language script (fe7adef)
• translate the two Chinese-language scripts, and keep source English (c1c904e)
• two of the day's three headline numbers do not survive re-derivation (99a348c)
• v1 was the quadruped's, but not for the reason this file gave first (1da05a5)
• val was three cameras, and R7 was only ever written about test (a846a48)
• what a per-class number stands on, and which half of product is settled (5d683da)
• what actually blocks removing the traversability head, counted (1ae3c02)
• what three seeds cost, what they bought, and why an mtime says nothing (840785a)
• withdraw the night half of the 0.35 person threshold (d6d8df1)

Build & packaging

• configure pyright, so Pylance stops burying findings in import noise (2940e6e)
• keep customer site drawings out of git (8c7eaa2)
• make a # noqa prove it still suppresses something (e3134e7)
• repair the ty config so the checker starts, and resync the lock (3b2d12b)
• ship type annotations and wire a checker to read them (136cd0b)
• the pre-commit ruff and CI's ruff disagreed about what "formatted" means (4b0fe8e)
• watch the pinned actions and dependencies with dependabot (3737f75)

CI

• cut releases with release-please, split across dev and main (4db1b54)
• install from the lockfile, and hold the type count to a ratchet (d1d7402)
• scripts/ gets a type ratchet, on its own baseline rather than src/'s (ecdf2ee)
• test the Python the maintainer actually runs (fbd4b3c)

---

This PR was generated with Release Please. See documentation.