Skip to content

chore: bump splat-transform to 3.3.0, playcanvas to 2.21.4 (+ decimate flag rewire, LOD NaN fix) - #92

Open
CodeByKeegan wants to merge 3 commits into
devfrom
claude/optimistic-hopper-frf4dw
Open

CodeByKeegan wants to merge 3 commits into
devfrom
claude/optimistic-hopper-frf4dw

Conversation

@CodeByKeegan

@CodeByKeegan CodeByKeegan commented Aug 19, 2026 •

Copy link
Copy Markdown
Owner

Bumps @playcanvas/splat-transform 3.2.0 → 3.3.0 and playcanvas 2.21.3 → 2.21.4, rewires the renamed decimation flags, and fixes a LOD regression that arrives with 3.3.0.

Why this PR and not the other bump PRs

Seven near-identical bump PRs are open (#81, #82, #83, #89, #90, #91, plus Dependabot #85). They were produced by the dependency-update routine firing daily instead of weekly against an unmerged PR. This one is the consolidation point; the rest should be closed as superseded.

Do not merge Dependabot #85 — it bumps package.json/package-lock only. Every uniform-algorithm run would emit --decimate-uniform against a CLI that no longer accepts it and fail with ERR_PARSE_ARGS_UNKNOWN_OPTION. The dependency bump and the flag rewrite must land atomically.

Release review (step 2b)

splat-transform 3.3.0 — three upstream PRs: #302 (decimate flag swap), #300 (per-leaf LOD approximation errors), #303 (dep updates).

  1. The decimate flags swapped meaning. In 3.2.0 -d/--decimate was adaptive and --decimate-uniform was uniform. In 3.3.0 -d/--decimate is uniform (and the CLI default) with a new --decimate-adaptive. Verified by running both CLIs. This is the only change to the --help surface — no new flags to wire.

    The algorithms themselves are frozen: 3.2.0 --decimate-uniform output is byte-identical to 3.3.0 --decimate, and 3.2.0 --decimate is byte-identical to 3.3.0 --decimate-adaptive. So this is a rename, not an algorithm change — and there is no alias in either direction, so a mismatched flag fails loudly.

    Splat Studio keeps adaptive as its own default, passing --decimate-adaptive explicitly. decimateAlgorithm is a field on our MCP schema and /api/convert body documented as defaulting to adaptive; absorbing an upstream default flip would silently change output for every existing caller and preset that omits the field. The divergence is recorded in server/commands.mjs.

  2. New LOD hard-failure (the regression this PR fixes). 3.3.0's LOD writer aborts on any non-finite gaussian:

    ✗ Error: LOD 0 gaussian 7 has a non-finite position; run --filter-nan to drop invalid gaussians before writing LODs
    

    Reproduced against both versions with the same input: 3.2.0 exit 0 and a full bundle; 3.3.0 exit 1 and no bundle. Plain (non-LOD) conversion of the same file still succeeds, so this is specific to the LOD writer.

    server/commands.mjs already emitted -N in both LOD builders, but nothing could set it — filterNaN was exposed only on the convert panel and the convert MCP tool, so a LOD bake had no way to act on the error it was told to fix. This PR adds the Filter NaN checkbox to the LOD panel, sends it from lod-panel.ts, and declares filterNaN on the build_lod MCP schema. The existing server plumbing was already correct, including argument order.

    Ordering matters and is now covered by a test: -N is an action applied to the preceding input, so -N placed before the input does not suppress the error. Both LOD builders already push it after each input (including every combine row).

  3. LOD bakes are slower. The new per-leaf error pass is unconditional with no opt-out: on a 400K-gaussian 2-level bake, chunking went 0.99s → 3.36s and end-to-end 3.9s → 7.2s. Peak memory flat. lod-meta.json gains lodErrors: true and per-leaf errors[]; format version stays 1 and per-chunk payloads are unchanged, so the counts readers are unaffected.

  4. Library export trap (no action needed today). decimateSource silently swapped from the adaptive to the uniform implementation; the old behavior is now decimateSourceAdaptive. Harmless while we only shell out to the CLI, but a trap for any future in-process import.

playcanvas 2.21.4 — a single fix: "Discard out of range depth in VSM shadow casters" (#9124), touching only two shadow-pass shader chunks. No JS/TS source, no public API, no typings. Splat Studio creates one directional light and never enables castShadows (repo-wide grep for castShadows/shadowType/SHADOW_VSM returns zero hits), and the added line is gated behind #if SHADOW_TYPE == VSM_*. No-op for us. Note package.json already declared ^2.21.3, so this is effectively a lockfile re-pin.

Verification

  • npm run typecheck — clean.
  • npm test — 72 passed, 0 failed, including the new check streamed LOD: NaN source fails without Filter NaN, bakes with it, which asserts the failure, the actionable error text, the successful bake with the flag, -N argument placement, and that build-meta.json records the choice.
  • New UI control verified visually, not just via green tests: npm run docs:capture regenerated the annotated screenshots and the Filter NaN checkbox renders natively in the LOD panel.
  • Docs refreshed — README version line, docs/USER_GUIDE.md (LOD section), docs/MCP_WORKFLOWS.md (build_lod), and the workflows skill in both .claude/ and .agents/.

Follow-ups (not in this PR)

  • Consider narrowing to ~3.3.0. Upstream has now demonstrated it changes flag semantics in a minor release, and for uniform we emit bare --decimate, which depends on 3.3's meaning.
  • Sanity-check LOD job timeouts on a large scene against the ~1.8× slowdown.

Authored autonomously by the weekly dependency-update routine.

splat-transform 3.3.0 reverted the decimation default to uniform and
renamed the adaptive flag to --decimate-adaptive (the old
--decimate-uniform flag was removed). Updated the server's flag
mapping, UI tooltips, docs, and e2e tests to match; Splat Studio's own
default stays adaptive.

playcanvas 2.21.4 is a shadow-rendering patch fix with no API changes.

Regenerated the annotated UI screenshots and refreshed the README
version line.
tests/mcp-e2e.mjs and mcp-server/tools/convert.mjs still referenced
the removed --decimate-uniform flag; CI caught it via npm run test:mcp.

Copy link
Copy Markdown
Owner Author

test, CodeQL, and both Analyze checks are green. The github-advanced-security check is red on both commits with the identical failure — GitHub's own Copilot code-scanning session errors before it even reads the diff (SessionModelError: 400 The requested model is not supported, model sweagent-capi:claude-opus-4.6). That's infra-side model routing on GitHub's end, not something in this repo's workflow files, and it isn't blocking mergeability (mergeable_state: clean). Not pushing changes to chase it.


Generated by Claude Code

…ussians)

splat-transform 3.3.0 makes the LOD writer refuse any source containing a
non-finite gaussian ("LOD 0 gaussian N has a non-finite position; run
--filter-nan to drop invalid gaussians before writing LODs"), where 3.2.0
wrote the bundle. server/commands.mjs already emits -N after each input in
both LOD builders, but nothing could set it: filterNaN was exposed only on
the convert panel and the convert MCP tool, so a LOD bake had no in-flow
remedy for the error it was told to fix.

Add the Filter NaN checkbox to the LOD panel, send it from lod-panel.ts, and
declare filterNaN on the build_lod MCP schema. -N is an ordered action, so it
must follow its input — the e2e check asserts that placement alongside the
fail-then-succeed pair.
@CodeByKeegan CodeByKeegan changed the title chore: bump splat-transform to 3.3.0 (+ flag rename fixed) chore: bump splat-transform to 3.3.0, playcanvas to 2.21.4 (+ decimate flag rewire, LOD NaN fix) Aug 20, 2026

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant