feat(engine): ContentLayer registry + FrameStep program — renderer-unification plan 02#420
Merged
Merged
Conversation
…end, RenderTargetSpec, ContentLayer, CompositeStep, FrameStep, RenderStrategy Verbatim from the renderer-unification spec's locked Core-concepts blocks (plan 02, Task 1). No consumers yet — the registry and executor land in later tasks. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Deploying with
|
| Status | Name | Latest Commit | Preview URL | Updated (UTC) |
|---|---|---|---|---|
| ✅ Deployment successful! View logs |
skymap | 277e687 | Commit Preview URL Branch Preview URL |
Jul 09 2026, 09:07 AM |
deriveSlabs instantiates the spec's two-row table per frame (near-field f64 row layerless until the zoom-to-earth fold; cosmological f32 row wraps today's vp). slabViewOf narrows a row to the per-step SlabView the executor threads into layer draws. ReadyFrameContext gains slabs, populated in deriveFrameContext beside vp so there is exactly one derivation per frame. f32→f64 widening is exact, so the COSMO round-trip is byte-equal and slabViewOf needs no special case. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…EAR0 provisional-semantics note Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…_LAYERS registry Mechanical conversion per the migration table: slab COSMO, target hdr, blend additive on every row; renderer refs move from the PassDeps bag to state.gpu.*; vp/viewport reads go through the per-step SlabView. The old encoders keep driving the frame through a transitional HDR_PASSES filter until the program flip; UI_PASSES converts next. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…deleted CONTENT_LAYERS now carries all fourteen rows (nine hdr additive + five swap over); UI_PASSES joins HDR_PASSES as a derived transitional filter. Pass/PassDeps types are gone — layers read state.gpu.* directly, so RenderFrameInput and RunFrameDeps shed their renderer fields and the deps bag no longer threads through renderFrame/runFrame/startLoop. startLoop keeps its fail-fast renderer-readiness check (an independent boot invariant, not deps forwarding), repointed at state.gpu.*. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ming slots frameProgram(tone) emits the five-step main program (flow compute, volume render, hdr render, hdr→swap tonemap composite, swap render); timedSlotsOf derives the GPU-timing slot list from the program + layer registry so slot names can never drift from what the frame actually runs. Dead code until the renderFrame flip wires it in. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…E table Runs a FrameStep program into one encoder: the step-kind switch is the frame's only switch. Render steps select layers by (target, slab), get one SlabView each, and honour the disabledPasses one-way override; composites gate on a per-frame touched set and delegate to the Compositor; first pass against a target clears with its table value, later passes load. merged vs perLayerTimed rides as data — the timing fork is a strategy, not two encoder functions. encodeFlowCompute slims to (encoder, state) reading its own gates; its args bag is gone. Dead code until the renderFrame flip. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…icate-name trap Two unrelated types shared the name EngineState: the engine's runtime state bag (108 importers) and the Redux 'engine' slice shape (3). The slice is now EngineSliceState so the names can never be confused at an import site again. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
renderFrame shrinks to focus-uniform write + executeFrame(frameProgram, CONTENT_LAYERS) + submit; strategy derives from timingService.enabled. scalarVolumeLayer joins the registry (volume target, before the hdr group); deriveVolumeLiveness is the ONE shared projection both volume layers' gates consume, dissolving the prepass/upsample double-gate that could drift. Timing slots now derive from the program (TIMED_SLOTS); TIMED_SLOT_NAMES deleted; DebugPanel reads the same derived list. The engine-ready seam narrows compositor alongside the other required GPU handles. Old encoders are no longer called — deleted next. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… frame recorder encodeHdrSingle/Split, encodeUiOverlay, encodeVolumePrepass and encodeVolumes are unreferenced since the program flip; their transitional HDR_PASSES/UI_PASSES registry views go with them (engine.ts's debug name list reads CONTENT_LAYERS directly). encodeVolumes' assertions live on in executeFrame/scalarVolumeLayer tests. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…t the deleted encodeUiOverlay Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
createRenderTargets holds the spec rows (hdr rgba16float/1, volume
rgba16float/3, swap swapFormat/1), allocates and resizes every offscreen
texture in one place, and serves views by id ('swap' throws — the
executor resolves it per frame). postProcess and volumeOffscreen
dissolve; the readiness gate's two checks become the one renderTargets
check; runFrame's hand-enumerated resize pair becomes a single resize
call; the volume scale divisor now lives only on the 'volume' row.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…chain format, always Every renderer factory now declares its colour target as an explicit targetFormat (positional or named per each factory's existing style); no factory reads GpuContext.format for its target. flowFieldRenderer's hdrFormat and instancedQuad's config format rename for uniformity, and the shared instancedQuad factory narrows from ctx to device — it only ever used .device. The registry target↔renderer-profile invariant is now legible at every construction site. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…PassDeps/Pass mention Fifty files of prose caught up with the registry+program reality: hdr/swap-target layer groups instead of HDR_PASSES/UI_PASSES, executeFrame program steps instead of the deleted encoders, <name>Layer filenames. Two draw-order claims corrected against the actual registry order. Review nits folded in: typed vi.fn house form in executeFrame tests, Readonly clear-value table, discriminating ctx-vs-target formats in the four UI-overlay format tests. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…act; grouping is (target, slab) The radar found ContentLayer.blend has no frame-path reader (the pipeline bakes the real blend state; the executor groups by target and slab). The field's docs now say so honestly, drawPick-style, and name the layer↔pipeline parity check as the guardrail once a target's layers stop agreeing on blend. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
First non-zero raw pick value in near→far slab order, mirroring the visible far→near OVER composite as a CPU fold: near content occludes far, so the nearest slab's hit wins. Plan-03 Task 1. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Derives a ready frame context from the last rendered pose (cameraRuntime.lastPose.current + projection) through the same side-effect-free deriveFrameContext math the frame uses, with the PICK source mask as visibleSourceMask. At rest this reproduces the frame's camera exactly — the prerequisite for deleting the pick uniform byte-stash. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Assembles the packed-field subset (selection sentinel, catalog + bias settings, LOD fade constants) and delegates to packPointUniforms, which stays the single byte-layout truth — drift between the pick and visual uniforms is structurally impossible. The packer's settings param narrows to the packed subset via Omit; the three draw-only fields were never read by it. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
hoverPickDriver (via a uniformBytes thunk on its deps), wireInput's click path, and drawPickDebugOverlay all derive their bytes through pickFrameContext → slabViewOf(COSMO) → pickUniformBytesOf instead of reading the per-frame byte stash. Null-guard shapes preserved (not-ready now skips exactly where null-bytes used to). The stash is now written but never read — deleted next. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
state.picking.lastFrameUniformBytes is gone: every consumer rebuilds the uniform at pick time from the slab view, so the stash was write-only. PointRenderer.draw sheds its buffer return (it existed only to feed the stash); the lastFrameCam snapshot now updates on every draw — the preserve-on-null rule lost its trigger with the bytes gone, and the unconditional stash is the more accurate input to the MW pick gate. Doc sweep rewords every stashed-bytes mention to built-at-pick-time. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…e surface Extracts the uniform upload, the three pick overrides, the group binds and the per-source loop from recordPickPass; pick()/renderForDebug() delegate to it ahead of the ring/disk/MW fold-ins, so behaviour is unchanged until the pickProgram cutover. drawPoints uploads and binds group(0) even with zero sources — the ring and MW pick pipelines read the camera through that prefix, pinned by test. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…a stash dissolved proceduralDiskRenderer's five cached camera fields (written by the visual draw, replayed by pickDisks) are gone; the pick uniform packs from caller-supplied values with an unchanged layout. The instance replay stays — that is content (the last-drawn disk LOD set), not camera. The interim pickRenderer fold-in unpacks the camera from the uniform bytes its callers already pass; that call site dies at the pickProgram cutover. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
pointSprites (drawPoints with mask-filtered sources + rebuilt uniform bytes — the group(0) prefix every other pick pipeline reads), proceduralDisks (pickDisks with the slab-view camera), milkyWay (pickMilkyWay; visibility rides the enabled gate against the pick camera) and structureMarkers (pickRing). Registry-driven only — no name list; the rows are dead until the pickProgram executes them. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Registry-driven (drawPick presence + enabled, grouped by slab, resolved near→far with frontmostPick + unpackPick), one encoder one submit, mapAsync-throttled with the ported inFlight/destroy guards. Owns its pick targets lazily per slab — pick:near0 never allocates at N=1. drawPick gains a documented postcondition: leave group(0) bound to the shared camera prefix — pickRenderer.bindCamera is the narrow restore surface and the disks row restores after its own slot-0 bind, keeping the program name-blind. Dead until the cutover. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…n pickProgram
hoverPickDriver slims to { state, pickProgram, store, resolveDeps } —
the collectTargets/viewport/pointSize/timing/uniformBytes thunks all
dissolve into the program. Click resolution takes bare texel coords;
the debug overlay calls renderForDebug() and keeps its post-frame
placement as a latency choice, not a data dependency. The old
pickRenderer.pick path is now uncalled by production and dies next.
Documented behaviour deltas (degenerate scenes only, per the plan):
1. All-hidden scene: an empty pick pass runs and decodes null — a click
now clears a stale selection where nothing dispatched before.
2. Zero-catalog scene with visible rings: click and hover now agree
(both pick); rings were hoverable but not clickable.
3. Pre-first-frame window: the no-bytes-yet gate becomes engine-not-
ready; lastPose seeding means the same pose frame 1 renders.
4. MW-only/zero-catalog: lastFrameCam updates every draw (was preserved
-on-null) — the MW pick gate answers for the actually-rendered pose;
a latent-bug fix.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
pick(), renderForDebug(), recordPickPass and its fold-in tail, the
readback textures/staging/inFlight machinery and the fold-in renderer
constructor params are gone — the pickProgram owns the pick lifecycle.
Surface: { label, drawPoints, bindCamera, destroy }. collectPickTargets
dies with them (mask filter lives in pointSprites' drawPick; the hasAny
rule in the program's registry filter). Deleted test coverage accounted:
ported to pickProgram/proceduralDiskRenderer suites or died with the
fold-in order they asserted.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…chain The pick cutover moved the MW pick gate onto milkyWayLayer.enabled with the pick-time context, leaving milkyWayPickVisible callerless and lastFrameCam — its only reader gone — a write-only mirror of the frame camera: the last vestige of the frame→pick coupling this phase deletes. Both go, along with PickFrameCam (single-consumer type). Draw and pick now share ONE gate, which is stronger than the old mirrored-pair can't-drift claim. EnginePickingState stays (pickInFlight + pointerDown remain live). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…he complete pick image The sentinel, the padded point size and the pick-pass flag were written twice: once into the bytes, again as post-upload overrides in drawPoints (the two docblocks already disagreed on the count). pickUniformBytesOf now packs all three (byte-identical to the old end-state, pinned by test); drawPoints uploads verbatim and sheds its pointSizePx param. PICK_PADDING_PX moves to a GPU-free data module. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… verified Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… the program, no dead-symbol references Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
/feature-done audit: tests 3938/3938 (623 files) PASS, typecheck PASS, all checkboxes ticked (user visual gates for both plans passed 2026-07-09), no new TODO markers, no unowned deferred items — follow-up register lives in PR #420 body + .superpowers/sdd/progress.md. Spec + 00-index stay in place for plan 04 (fold zoom-to-earth on the #386 branch). picking-gpu-subsystem backlog item stays — orthogonal GPU-resource-ownership migration, out of this spec's scope. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Executes renderer-unification plan 02 (registry + program) AND plan 03 (pick fold-in) via subagent-driven development, on one branch per user request. 29 commits, each task individually reviewed; final whole-branch review in progress.
Replaces the two-registry frame (HDR_PASSES / UI_PASSES + bespoke volume prepass) with one flat ContentLayer registry drawn by a FrameStep data program through a strategy-parameterized executor; then folds picking in as a drawPick aspect executed by a parallel per-slab pickProgram with a pure CPU frontmostPick resolver. Net effect: frame order and pick order are DATA; both byte-stash camera couplings are deleted.
Plan 02 — COMPLETE (pending visual gate)
All 11 tasks + interstitial EngineSliceState rename. Entanglement radar: all 8 locked axes survived; 1 doc finding fixed (blend documented as forward contract). Suite green throughout.
Plan 03 — COMPLETE (pending visual gate)
All 12 tasks. Radar: all three named braids verified dissolved (lastFrameUniformBytes AND lastFrameCam extinct; camera travels as values; no frame-ordering coupling); F2 residue (pick-override double-home) fixed behaviour-neutrally; the dead milkyWayPickVisible/lastFrameCam chain deleted.
Documented behaviour deltas (degenerate scenes only — plan §Documented behaviour deltas)
?gpuTimings— DebugPanel rows under the new slot names (scalar-volume, per-layer hdr rows, hdr→swap, per-layer swap rows, pick). Frame visually identical to main.?showPickBufferoverlay matches the scene while orbiting.Follow-ups filed (deliberately NOT in this behaviour-neutral PR)
Note: the plan-03 "shader diff empty" parity check shows two comment-only hunks (sweep rewording); zero functional WESL changes, selectionEncoding parity test untouched.
🤖 Generated with Claude Code