Interface LODGroupRegistryDeps

Whether the owning loader has latched an archive fault.

Whether a loader under this LOD group has a recorded network failure.

Resident-byte budget (the ceiling). The single, adaptive VRAM budget shared with the GPU buffer pool — one authority, not a competing one. Omitted ⇒ no eviction (pure retention), the default for unit tests that construct the registry directly.

Measured total resident VRAM bytes (active + pooled) from the GPU buffer pool — the single accounting truth. The registry compares this to the budget to decide when to demote cold levels; it no longer keeps its own per-level byte estimate. Omitted ⇒ no eviction (pure retention), the default for unit tests.

Current view-update version (SceneLoader.currentViewVersion). Used for the slice-aware fallback: a gsplats level whose committed geometry was stamped with an older version still shows a previous slice/displayDims, so the registry treats it as stale and displays the coarsest level that IS fresh until the re-slice commits. Omitted ⇒ no freshness tracking (every ready level is treated as fresh — identical to the pre-feature behaviour; the default for unit tests that don't exercise scrubbing).

Monotonic wall clock in milliseconds, for the stale-hold budget (see config.lod.staleHoldMs). Injectable so tests can advance it deterministically; omitted ⇒ performance.now().

Keep the render loop alive (the viewer is on-demand and idles after ~2s). Called each frame while a lazy level is loading so a deferred fine reload — which commits OUTSIDE the per-slice sweep and can take longer than the idle timeout — still triggers the per-frame swap-up to the fresh level when it lands, instead of waiting for the next user interaction. Wired to AnimationController.startAnimation (resets the idle timeout). Omitted ⇒ no-op (unit tests don't run a loop).

Keep the loop TICKING without asking for a redraw — the render-on-change counterpart of requestRender for the two keep-alive calls above (a lazy level loading, a pending partition resync): nothing drawn has changed yet, the loop just has to keep running the per-frame selector so the swap fires when the load lands (a swap marks the frame dirty through LODGroupRegistry.takeDrawnStateChanged). Wired to AnimationController.requestTick. Omitted ⇒ those calls fall back to requestRender (an embedder whose host renders on every wake).

Re-run the current view update for the partition parts that just re-entered the frustum. Culled children skip event-driven slice updates, so the rising edge must resync any geometry that missed the latest view state. paths are the re-entering PART node paths (child.path), or the partition wrapper path when a part has none (= resync the whole partition); the loader prefix-matches its sweep loaders under them, so a nested kind=lod part's eager level is covered. The view state is unchanged, so the loader MUST NOT bump the view version for this pass — otherwise every lazy fine level scene-wide reads stale and drops to coarse.

The view the drawn geometry was last COMMITTED for (the owning loader's last committed pass). Partition parts whose bounds miss its hidden-dim slice draw nothing, so they are hidden and kept out of refinement (B4) — against the COMMITTED view, not the requested one, so a part leaving the slice disappears in the same frame its successors are committed rather than when the pass that replaces it starts. Omitted / undefined ⇒ no slice gating (every part is in slice — the pre-B4 behaviour).

Whether the owning loader has a view PASS in flight or queued. Rising edges are held (and coalesced) while this is true so a resync never lands on top of a pass. A refinement hold deliberately does NOT count: the loader parks a resync that arrives during one and cancels into its own pass, so re-entering parts do not sit on a stale slice until the ladders finish.

Whether the LOD level dissolve is enabled (ON by default; ?noLodFade disables; the name predates the time-driven dissolve). When true and a blendable (additive/luminous/volumetric — see BLENDABLE_MODES in scene/lod-fade.ts) group changes its displayed level, the registry dissolves from the outgoing level to the incoming one over config.lod.fadeMs — the incoming at smoothstep(elapsed / fadeMs), the outgoing at the complement — instead of a hard visibility swap. Time-driven, so a parked camera always settles on ONE level, and independent of additive streaming. Omitted / false ⇒ the hard swap, byte-identical. Read live so the flag applies without a reload. The default for unit tests (off).

The period (ms) of the fastest dimension currently PLAYING, or null when nothing plays (DimensionAnimationManager.getPlaybackPeriodMs). While non-null, lazy levels reload on every timepoint without the settle debounce and the aspiration is capped at the finest level whose measured load time fits (see config.lod.playbackLoadBudgetFraction). Omitted ⇒ never playing — identical to the pre-feature behaviour.

Whether streaming brightness compensation is enabled: as a blendable (additive/luminous/volumetric) leaf's ladder streams in, scale its opacity by 1/e(k) so the partial prefix renders at the full-level energy (no brightening pop). Distinct from the level dissolve (within one level's stream, not between levels) and independently gated; either flag on enables the registry's per-frame opacity management. Omitted / false ⇒ byte-identical (no material writes). Read live so the flag applies without a reload. The default for unit tests (off).

Force the finest LOD level regardless of projected screen coverage (and never coarsen off-screen) — for high-quality still/video capture (the gallery harness), where a coarse level looks blurry even when the subject is small in frame. Wired from LuxarAppOptions.lodFinest (the ?lodFinest URL flag, threaded through the standalone bootstrap); omitted / false ⇒ normal coverage-driven selection. Read live, like the sibling flags above.

Replacement-LOD bias in screen-area units. 2 advances one level on an occupancy-halved ladder. Legacy diagonal coverage receives sqrt(bias) so both selectors shift by the same area factor. Since finite screen-area coverage is at most 1, bias below 1 makes a partition-anchored finest level unreachable and bias below 0.5 does the same for a whole-object finest level. Invalid values are neutral.

Register a clone-on-first-fade material with the material manager so it keeps receiving per-frame camera-uniform updates (the fade clones the shared cached material to fade one level independently; an unregistered gsplat clone would project with stale camera params). Wired to materialManager.register; omitted in unit tests (no camera loop).