CHANGELOG.md

Changelog

The format is based on Keep a Changelog.

Rearchitected at 0.4.0. Through v0.3.0 the extension ran an in-Omniverse matching engine. Starting with 0.4.0 it is a lightweight client for the hosted Physna Scan Search API — the two designs share almost no surface. The pre-0.4.0 local-matching history is preserved in git (tags v0.1.0–v0.3.0) but is not carried forward here.

[0.6.0] - 2026-08-04

Open-source release. This version prepares the repository for its public home: proper licensing, a support disclaimer, cleaned-up docs, and the removal of the last dead code left over from the pre-0.4.0 in-process matching engine.

Added

• Apache License 2.0. A root LICENSE file, Physna SPDX headers on every source file, and the license shipped inside the release zip.
• Support disclaimer (docs/DISCLAIMER.md, also shown on the extension's documentation pages and summarized in the README): the project is open source and maintained best-effort — large issues and failures are prioritized; minor issues and feature requests may not receive a timely response. The hosted Physna platform remains a separate commercial service.

Changed

• Default environment is now Production (app-api.physna.com) instead of the internal dev stack; a fresh install's Environment dropdown starts on "Production (app)" with the matching token endpoint pre-filled.
• Developer docs (docs/DEVELOPING.md, AGENTS.md) now use generic Python-environment instructions instead of machine-specific paths.

Removed

• PointCloudConverter — dead code from the in-process matching era (USD/STL→NPZ export, Open3D mesh reconstruction); nothing called it. CAD parts still convert to USD for placement via MeshConverter.
• The NVIDIA-proprietary Kit-template license headers (replaced by the Apache-2.0 headers above).
• The internal API guide (docs/API-Guide-ScanSearch.md) — it carried internal stack URLs and a tenant ID and had drifted out of date; the workflow summary lives in docs/Overview.md.
• Stale editor/tooling config: tracked .vscode/ files (already gitignored) and an unused renovate.json.

[0.5.0] - 2026-07-16

Reliability + workflow release: interrupted searches are now resumable, placed parts land at real-world size, E57 reads can no longer crash Kit, and the panel got a broad UX/performance pass.

Added

• Resume interrupted searches. Runs are checkpointed to disk right after upload (marked incomplete) and finalized on completion — if Kit closes mid-search, the run reappears in Previous Searches with a Resume button that re-polls to terminal and reads matches. Refresh (and startup discovery) also reconciles: an interrupted run that finished on the platform while Kit was closed flips to complete automatically.
• Placement size normalization. Physna's match transform is rigid (no scale), so parts are now auto-scaled to real-world metres: USD parts by their authored metersPerUnit, converted CAD parts assuming millimetres.
• Scene up-axis control. File-loaded scans are rotated to the stage's up-axis (Z-up assumed, flippable to Y-up / as-is); placements rotate with the scene prim.
• Environment dropdown (Production / Dev3 / Dev2 / Custom) fills API Base
  • Token URL together; the URLs moved under Advanced.
• Multi-select part picking (Ctrl/Shift-click in Add File(s)) and remembered last-used directories per picker.
• Matches list pagination + filter, same pattern as Previous Searches.
• "Scene was a stage prim" chooser: when a loaded run's scene prim isn't in the stage, pick a local file or download the uploaded scan back from the platform (cached on the record for later loads).
• Button tooltips (custom-drawn with explicit colors — Kit's default tooltip style is unreadable in this panel), two-click Delete ("Confirm?"), per-second poll countdowns, and add-part feedback.
• Pure-Python unit tests for the run store, API state model, deps gate, and path helpers (physna/reality_compiler/tests/).

Changed

• E57 files are read in an isolated subprocess — pye57 segfaults inside Kit's process (native DLL conflict), so a crash is now a reported error instead of taking Kit down. E57 is no longer offered in the scene file picker (raw .e57 uploads don't match on the platform): import via File > Import and use "Use Selected Prim".
• Pip installs no longer block the UI: only requests/keyring install synchronously; the point-cloud format libraries install on a background thread, and a file load that races the install waits for it (with status).
• Dragging a part's Placed slider no longer rebuilds the Matches section (or steals scroll); only the placement-dependent buttons refresh, and only when the any-placed state flips.
• Long operations (search, load, resume, place/clear all, downloads) are serialized — one at a time, each with an inline progress row.
• Asset downloads stream to disk (no full-file RAM buffering) with throttled progress; scan-match reads retry through the platform's transient "not been computed" 404 instead of surfacing it as an error.
• All temp scratch consolidated under <os-temp>/physna_reality_compiler/; stale isolated-E57 output is swept on the next read.

Fixed

• Re-running a search no longer leaves the previous run's placements in the stage attributed to the new results.
• Saved-search records keep their original creation time on update/resume (the list no longer reorders on every save) and are written atomically.
• A queued duplicate of the same part file can't collide on the upload folder; per-part scene-match read failures are keyed by asset id.
• An interrupted run whose assets were deleted on the platform no longer lingers in Previous Searches as resumable forever: Refresh detects the deletion (asset 404s are no longer retried indefinitely) and removes the stale local record. Update/Resume on a deleted run reports one clear "deleted from the platform" error instead of polling to timeout or spraying raw per-part 404 bodies, and a part deleted individually gets a short human-readable note while its locally saved matches stay loadable.
• File-picker dialogs are destroyed after use (one window leaked per pick).
• Sign in / Refresh double-clicks, mid-run slider drags, and mid-run up-axis changes are guarded; extension unload cancels in-flight tasks.

[0.4.0] - 2026-07-14

Complete rearchitecture: the extension is now a thin client over the hosted Physna Scan Search API. All point-cloud matching, registration, and scoring runs on the platform; the extension handles selection, upload, polling, placement, and visualization inside Omniverse.

Added

• physna.reality_compiler.api — pure-Python client for the hosted API (no omni/pxr imports, so it is testable standalone):
  • OAuth2 client-credentials auth with an in-memory token cache
  • secure service-account storage via the OS credential vault (keyring)
  • upload_asset / get_asset / get_scene_matches / delete_asset
  • asset-state model + terminal/queryable predicates and batch polling
  • ApiSession login/logout/restore facade
• pipelines/workflow.py — ScanSearchWorkflow: prepare an uploadable scene (picked file or points extracted from a selected prim), upload scene + parts into a unique colocated run folder, poll to a terminal state, read scene-matches, and place matched part USDs at the returned 4×4 transform. Per-part matches stream in the moment each part (and the scene) finish indexing, instead of after the whole batch.
• Native accordion property-panel UI (Account / Scene / Parts / Run / Results) styled to match Kit's Property panel.
• Results: a per-part match slider (drag to place/remove matches live), combined Matches + Placement with a Min score slider that gates how many matches place, and Place All (min score).
• Placement hot-swap: scan points behind placed matches are hidden (reversibly, via per-point widths) so parts aren't occluded — on by default.
• Previous Searches: local and platform runs in one list with a debounced live filter, pagination, and inline per-row download progress; discover and download runs created anywhere in the tenant (folder layout preserved so USD references resolve).
• Point-cloud loaders: NPY/NPZ, ASCII XYZ/PTS, E57 (pye57), LAS/ LAZ (laspy + lazrs), PLY (trimesh), and PCD (pypcd4); non-native scans are uploaded as extracted .npy.

Changed

• extension.toml: dependencies trimmed to the Kit surface the client needs; pip deps (requests, keyring, laspy, lazrs, pye57, trimesh, pypcd4) are declared via [[python.pipapi.install]] and also installed by extension.py at import time as a runtime backstop.
• PipelineManager is now an API-session + orchestration facade.
• Placement is batched (one Hydra sync per set) and instanceable (placements of the same part share one GPU prototype); point removal runs off-thread — so importing many matches and carving the scan no longer hitches the UI.

Removed

• The in-process matching engine (scansearch/: GPU kernels, registration, scoring, matching, nearest-neighbour, parallel batch).
• docker_bridge/ (bridge to the old self-hosted matcher container).
• Local search-tuning UI, plane-removal previews, and batch/staged multi-class orchestration (all handled server-side now).
• Heavy runtime dependencies: open3d (PLY/PCD now load via trimesh/pypcd4), scipy, networkx, scikit-learn, and omni.warp.core.

Security

• NPZ metadata loads with allow_pickle=False (closes a pickle code-execution path through a crafted .npz); converters write JSON sidecar metadata instead of a pickled object array.
• The service-account secret is never sent over a non-HTTPS connection (loopback http is allowed for local dev), is redacted from repr, and is scrubbed from API error bodies.
• A load-size cap (2 GB) guards against a hostile or corrupt "point cloud" exhausting memory before a byte is read.

Fixed

• color.npy / segmentation.npy sidecars load again — a scan's per-point color was being dropped — now read with allow_pickle=False.
• A part that finishes indexing early shows its matches immediately rather than reporting "0 results" until the slowest part completes.