Write-back coverage and gaps: measured inventory

Recorded on 2026-10-08 to answer what DFX can and cannot write back, and which editor changes cannot be synchronized back to text.

Scope: read original assets and write source text, with the embedded-script extraction exception described in section 3.2. Numbers come from one continuous local run on a UE 5.8.2 source engine with DreamFX 359b3b3, covering all 59 Niagara systems under /AtlasFX, /Game/GameActor2D and /Game/AssetMaterial. Commands and artifacts are listed in section 7. Anything not exercised is explicitly unverified. Later implementation results are dated separately in sections 6.8–6.10; this is a historical measurement record, not a claim that every initial gap remains in current code.

The separate research note write-back-from-dreamgui.md covers DreamGUI/DUI's approach and techniques DreamFX could reuse. This document records DFX's measured gaps, without repeating that design discussion. The research note is not part of this documentation checkout.

0. Conclusions

Successfully reading an asset does not prove it can be written back.

• Read: 59/59 systems exported; only six assets reported 17 unexpressible features. The coverage command's own reported expressibility was 100%.
• Write: rebuilding five samples from those 59 succeeded for four, while one could not even reach compilation (NS_ScanScene, 21 DFX4007 errors).
• The four successfully built samples passed L1/L2, but independent asset-reflection comparison with asset-diff found three assets with real differences, including a confirmed stored-value divergence (section 3.5). Its visual/runtime consequence was not verified.
• The specific old failure “mesh renderer material-parameter attribute bindings disappear and render a white block” did not reproduce in the measured export/rebuild path (reflection evidence in section 2.1). The same kind of loss occurred elsewhere: module-input literals, invisible to mirror-diff.
• mirror-diff (L1+L2) cannot alone establish lossless write-back. Include asset-diff, which itself had two coverage gaps (section 3.9). At the time of the initial measurement, bidirectional editing still needed an asset-level safety gate that sees state absent from text, and a supported read path for suppressed inputs. The only existing switch, -NoDefaults, was documented as diagnostic output not intended for maintenance.

1. Coverage check (coverage)

pwsh -File .skill/dfx.ps1 coverage -Path '/AtlasFX+/Game/GameActor2D+/Game/AssetMaterial' -Engine D:\UnrealEngine-5.8.2

59 systems scanned, 59 exported, zero failed. Reported gaps: six assets, 17 occurrences. Asset and emitter identifiers below retain their original spelling.

Three qualifications affect the conclusion:

1. coverage counts FDecompileResult::UnsupportedFeatures (Source/DreamFXEditor/Private/Commandlet/DreamFXCommandlet.cpp:309-381), the features the decompiler itself acknowledges it cannot express. It does not test generation. NS_ScanScene reports ok (2 gap(s)) here, but its export cannot rebuild at all (section 3.3). This measures read coverage, not write-back coverage.
2. Inheritance gaps are the documented DFX8014 flattening of the parent chain: the link is lost, not the contents (Docs/diagnostics/DFX8xxx.md:223-239). The measurement confirms this: all four inherited emitter stacks in TR刀光_System survive, while parent becomes none (section 3.1). The table accounts for 15 inheritance occurrences, including the three thrust emitters; the earlier 12-occurrence subtotal covered the four slash emitters in three assets.
3. These 59 systems contained zero Stage blocks, zero cases with more than one event handler, zero custom stage classes and zero RibbonRenderer blocks. DFX8015/DFX8016 were never triggered. They are implemented gaps unverified by this content, not confirmed absent.

2. Decompilation evidence

pwsh -File .skill/dfx.ps1 decompile-all -Path '/AtlasFX+/Game/GameActor2D+/Game/AssetMaterial' -Engine D:\UnrealEngine-5.8.2
# 59 written, 0 failed, under C:\CrossingVoid\DFX\Decompiled\**

Static inventory of 59 .dfs exports; counts are matching lines / matching files:

Every export header contains // Inline arithmetic is not recovered: expressions come back as equivalent hlsl { } blocks. This is a comment, not an expression. A text search for hlsl { matches all 59 files and must not be counted as expression degradation.

2.1 Material-parameter attribute bindings: the old gap is closed for this path

The manually authored system /Game/GameActor2D/Misaka/Material/Sk1/NS_Effects1_Mesh, with no text source and representative of the old white-block failure, exports this renderer:

// DFX/Decompiled/Game/GameActor2D/Misaka/Material/Sk1/NS_Effects1_Mesh.dfs:90-100
MeshRenderer
{
    bOverrideMaterials = true;
    bSubImageBlend = false;
    MaterialParameters = "{\"attributeBindings\":[{\"materialParameterName\":\"Sheet\",\"niagaraVariable\":{\"name\":\"User.Atlas\",…},\"niagaraChildVariable\":{\"name\":\"ResolvedTexture\",…}}],…}";   // :94
    Meshes = "[{\"meshParameterBinding\":{…},\"mesh\":\"/Script/Engine.StaticMesh'/ZDBridge/FX/FXDefault.FXDefault'\",…}]";   // :95
    OverrideMaterials = "[{\"explicitMat\":\"/Script/Engine.Material'/AtlasFX/M_FXAtlasSheet.M_FXAtlasSheet'\",…}]";          // :96
    SortMode = ViewDepth;

    Bind CustomSorting -> Particles.NormalizedAge;   // :99
}

Sheet ← User.Atlas.ResolvedTexture is in the MaterialParameters blob. R4 decompilation preserves structured property values without dedicated syntax as JSON strings: TryWriteJsonBlob in Source/DreamFXEditor/Private/Decompiler/DreamFXDecompiler.cpp:1009-1044, called at :1231-1237 and :1249-1254. Generation symmetrically inserts them as JSON (Source/DreamFXEditor/Private/Generation/DreamFXGenerator.cpp:276-295).

Reflection evidence, independent of text export: asset-diff reads properties directly with ExportText_InContainer (Source/DreamFXEditor/Private/Diff/DreamFXAssetFacts.cpp:425-436):

emitter Atlas2D_Mesh renderer 0:NiagaraMeshRendererProperties MaterialParameters = (AttributeBindings=((MaterialParameterName="Sheet",NiagaraVariable=(Name="User.Atlas",TypeDefHandle=(RegisteredTypeIn…
emitter Atlas2D_Mesh renderer 0:NiagaraMeshRendererProperties MaterialParamValidMask = 4095

The original and mirror match character for character in Saved/DreamFX/NS_Effects1_Mesh.{original,mirror}.facts.

The old failure does not reproduce on this engine/patch combination through this path. That specific gap can be closed, but the class of loss remains (section 3.4).

2.2 Bind and MaterialParameters are different mechanisms

• Bind X -> Y; handles fields of FNiagaraVariableAttributeBinding, such as PositionBinding, ColorBinding and CustomSortingBinding. Reading visits only FStructProperty values of exactly that struct type (Source/DreamFXEditor/Private/Adapter/DreamFXNiagaraAdapter.cpp:2662-2675). The decompiler deliberately skips property keys ending in Binding, since Bind lines represent them (Source/DreamFXEditor/Private/Decompiler/DreamFXDecompiler.cpp:1178-1183).
• MaterialParameters is a TArray<FNiagaraMaterialAttributeBinding>, a different struct, using the JSON-blob path rather than Bind statements.

This matters for write-back: MaterialParameters has no strongly typed write path equivalent to Bind. Its measured round trip succeeds (section 5), but validation depends on FJsonSerializer and the engine's JSON-to-struct converter. Conversion errors are engine errors, not DFX-specific diagnostics.

2.3 Meshes and OverrideMaterials did not export as the documented path arrays

Docs/language/dfs.md:302-316 shows:

Meshes            = ["/Engine/BasicShapes/Cube"];
OverrideMaterials = ["Plugin.MoonToon:Materials/FX/M_Chunk"];

Measured output: 90 Meshes and 80 OverrideMaterials occurrences, all JSON blobs, zero path arrays. TryWriteReferenceArray (DreamFXDecompiler.cpp:1055-1133) requires both successful GetArrayElementReferenceField and all other element fields equal to a default-constructed element. Real elements carry export-text references such as /Script/Engine.StaticMesh'…' in mesh/explicitMat; TryReadReferenceObject (:920) cannot read these, so output falls back to the blob branch (:1249-1254).

No data is lost: blobs are lossless and all five available mirrors passed L1. However, the documented authoring form differs from actual export output. Documentation and editor integrations must not assume exports contain path arrays.

3. Unsupported write-back and loss inventory

Status after the second pass on 2026-10-09: sections 3.1/3.2 remain declared gaps; 3.3's namespace asymmetry remains; 3.4's value write-back is addressed by pull (-NoDefaults supplies keys, the formal store reader supplies values); 3.5 is addressed as described in 6.9; 3.6/3.7/3.8 were unchanged. Sections 3.9/3.10/3.11 describe tool properties and measurement boundaries. Section 6.10 records later disposition item by item.

3.1 Parent chains (declared, DFX8014)

The link, not the content, is lost. TR刀光_System, with four reported gaps, receives this L1 verdict: bodies identical; the original records 4 gap(s) its mirror does not have.

3.2 Unresolvable module scripts (declared)

In /Game/GameActor2D/Origin_Miku/SceneScan/Niagara/NS_ScanScene, GetNormalizedDistance and UpdateScanSpriteSize report has no resolvable script asset (scratch pad or missing reference). R3 in decompile-all extracts them as independent script assets:

extracted embedded script 'GetNormalizedDistance'   -> /Game/Decompiled/GameActor2D/Origin_Miku/SceneScan/Niagara/Scripts/NS_ScanScene_GetNormalizedDistance
extracted embedded script 'UpdateScanSpriteSize'    -> /Game/Decompiled/GameActor2D/Origin_Miku/SceneScan/Niagara/Scripts/NS_ScanScene_UpdateScanSpriteSize

Side effect: decompile-all does not only write text. This step writes two .uasset files under /Game/Decompiled/…/Scripts/. Content/Decompiled/ is ignored by .gitignore. Any description of this command as read-only must state this exception.

3.3 An entire exported document cannot rebuild (undeclared)

/Game/GameActor2D/Origin_Miku/SceneScan/Niagara/NS_ScanScene was the only sampled system whose build failed directly: 21 errors, all of the same kind.

error DFX4007: Input 'Divide_Float.A' expects NiagaraFloat. 'NPC.NPC_Scan001.MaxScanRange'
  is neither a literal of that type nor a parameter reference -- parameter references start
  with a namespace such as User., Particles., Emitter., System. or Engine.

This is the strongest read/write mismatch in the initial run: the reader emits a token the same plugin's writer rejects, while coverage reports ok.

3.4 Module-input literals absent from the source (undeclared)

This is the same class of loss as the old white-block issue, at a different location. The decompiler's default-suppression rule (DreamFXDecompiler.cpp:1730-1731) is:

R8: only inputs that differ from a pristine instance of the same module are printed.

More values than expected were stored in the original's RapidIterationParameters but absent from text. Reflection-level fact-set comparison with asset-diff -DumpFacts found:

Examples from Saved/DreamFX/破空灰尘.original.facts, all absent from the mirror:

ri Fountain ParticleSpawnScriptInterpolated Constants.Fountain.InitializeParticle.Sprite Rotation Angle (NiagaraFloat) = 0000F041
ri Fountain ParticleSpawnScriptInterpolated Constants.Fountain.InitializeParticle.Sprite Size (Vector2f) = 0000204100002041
ri Fountain ParticleSpawnScriptInterpolated Constants.Fountain.AddVelocity.Velocity Speed (NiagaraFloat) = 0000C841
ri system-update Constants.Fountain.EmitterState.MaxDistance (NiagaraFloat) = 00409C45

Interpret these counts carefully:

• Suppression is intended to be behaviorally equivalent: module defaults should restore the omitted values on rebuild. These facts alone do not prove incorrect rendering.
• They are nevertheless absent from this text. For source-authoritative editing this means:
  1. No supported read representation: normal decompile output has neither the input value nor a marker saying a value was suppressed. Only -NoDefaults reveals it; at the time, .skill/dfx.ps1:127-129 explicitly said “Diagnostic only -- the result is not meant to be maintained.”
  2. No write-back address: line-based write-back and a dirty set need the input's key, which is not present in the source at all.
• It is fragile across module updates: if module defaults change, the mirror follows them, while the original retains its stored values.

3.5 Confirmed stored-value divergence: MeshYaw −90 → 0 (undeclared)

/Game/GameActor2D/Misaka/Material/Sk1/NS_Effects1_Mesh has the smallest of the three measured asset differences: 136 fact(s) vs 136, two pairs, representing the same input in spawn and update scripts.

ORIG  ri Atlas2D_Mesh ParticleUpdateScript Constants.Atlas2D_Mesh.Sprite_Atlas_Size.MeshYaw (NiagaraFloat) = 0000B4C2
MIRR  ri Atlas2D_Mesh ParticleUpdateScript Constants.Atlas2D_Mesh.Sprite_Atlas_Size.MeshYaw (NiagaraFloat) = 00000000

0xC2B40000 is −90.0; 0x00000000 is 0.0. These are different values for the same variable. Every link in the following chain was measured:

1. The source omits MeshYaw. NS_Effects1_Mesh.dfs:87 contains only /AtlasFX/Modules/Sprite_Atlas_Size(DepthScale = User.DepthScale);.

2. The original stores −90, as shown above.

3. -NoDefaults, which prints all inputs, nevertheless exports 0:

   // Saved/dfx-writeback-study/NS_Effects1_Mesh.NoDefaults.dfs
   /AtlasFX/Modules/Sprite_Atlas_Size(
       MeshSize = (92.800003, 64.0),
       MinSize = (92.800003, 64.0),
       DepthScale = User.DepthScale,
       UniformScale = 1.0,
       MeshYaw = 0.0        // The asset's rapid-iteration store contains -90.
   );

   The same export of /AtlasFX/Templates/NS_Atlas2D_Mesh gives identical MeshYaw = 0.0. The read path therefore cannot distinguish “really zero” from “stored as −90.”

4. Across both templates and the Misaka system, only this input differs in value; MeshSize, MinSize and UniformScale agree on both sides.

Confirmed: source does not completely describe the original module-input state; the reader's MeshYaw differs from the rapid-iteration store, and the mirror differs from the original. Without -NoCompile, asset-diff forces both sides to recompile and waits (DreamFXCommandlet.cpp:816-827), so this is not merely a stale, uncompiled value.

Unverified: whether Niagara runtime consumes the pin or rapid-iteration value, and therefore whether the mirror actually renders differently. That needs L3 or visual comparison.

Related context, not proof: an AtlasFX source comment in Plugins/AtlasFX/Source/AtlasFXEditor/Private/AtlasFXPaperZDPreview.cpp:63 attributes an earlier invisible preview to mesh-template MeshYaw = -90, which rotated the plane edge-on to the camera (also a thin line in game); the template was changed back to zero.

This was the only initial clue connecting a definite source/asset value difference with a possible visible change, and merited investigation first.

3.6 Renderer MaterialParamValidMask 15 → 0 (undeclared)

In emitter 位置, one of four mesh renderers in /Game/GameActor2D/SAO_Kirito/ExAsset/TR刀光_System:

ORIG  emitter 位置 renderer 0:NiagaraMeshRendererProperties MaterialParamValidMask = 15
MIRR  emitter 位置 renderer 0:NiagaraMeshRendererProperties MaterialParamValidMask = 0

The same asset also loses a group of dynamic-material-parameter RI constants:

ORIG  ri 位置 ParticleSpawnScriptInterpolated Constants.位置.DynamicMaterialParameters.Index 0 Param 2 (NiagaraFloat) = 00000000
ORIG  ri 位置 ParticleSpawnScriptInterpolated Constants.位置.DynamicMaterialParameters.Index 0 Param 3 (NiagaraFloat) = 00000000
ORIG  ri 位置 ParticleSpawnScriptInterpolated Constants.位置.DynamicMaterialParameters.Index 0 Param 4 (NiagaraFloat) = 00000000

The source does contain DynamicMaterialParameters@1.1(…) at TR刀光_System.dfs:528. The module exists; the missing state is the renderer's valid mask and corresponding RI constants.

Unverified: whether MaterialParamValidMask is a cache recalculated during load/compile. It is not CPF_Transient (otherwise AppendPropertyFacts, DreamFXAssetFacts.cpp:85-89, would skip it), so it is serialized asset state. Whether recompilation restores it was not separately tested. Its consequence remained unverified, alongside section 3.5.

3.7 Curve/DI node-name drift (undeclared)

The compiled … di '…' fact family differs in 破空灰尘 and TR刀光_System:

ORIG   compiled Fountain  ParticleSpawnScriptInterpolated di 'Emitter.Scale Alpha.FloatCurve'    … calls: SampleCurve
ORIG   compiled Fountain001 ParticleSpawnScriptInterpolated di 'Emitter.FloatFromCurve005.FloatCurve' … calls: SampleCurve
MIRR   compiled Fountain  ParticleSpawnScriptInterpolated di 'Emitter.FloatFromCurve.FloatCurve'  … calls: SampleCurve
MIRR   compiled Fountain001 ParticleSpawnScriptInterpolated di 'Emitter.FloatFromCurve004.FloatCurve' … calls: SampleCurve
MIRR   compiled 刀光纹理   ParticleSpawnScriptInterpolated di 'Emitter.FloatFromCurve006.FloatCurve' … calls: SampleCurve

Both sides call SampleCurve, suggesting similar behavior, but this is not a behavioral proof. The curve-owning module node names drift (Scale Alpha → FloatFromCurve, FloatCurve001 → FloatFromCurve006). as <name> in dfs.md:120-126 addresses stable node identity, yet none of the 59 exports contains an as pin.

Unverified: whether this can break Output.<node>.<value> links. No resulting compile error appeared in the available L2 checks, but the absence of as is established.

3.8 EffectType: asymmetric read/write tables (code finding, not exercised)

The comment at DreamFXDecompiler.cpp:1266-1274 requires the decompiler table to stay aligned with generator EmitterSettings: the two are halves of one contract, and a setting present on only one side is silently lost. Inspection found:

At this revision, EffectType can be written but not read back. Export/rebuild can silently omit it for systems that use it. This content had zero EffectType = occurrences, so the failure was not exercised. The asymmetry itself was real and could be corrected with one table entry.

3.9 Verification-tool blind spots

3.10 Known normalization noise

These differences may come from the comparator rather than lost authoring state:

1. <self> prefixes: original emitter Fountain renderer 0:… Meshes = ((Mesh="/Script/Engine.StaticMesh'<self>_Mesh.破空灰尘_Mesh'")) versus mirror …'StaticMesh'/Game/AssetMaterial/FXs/破空/破空灰尘_Mesh.破空灰尘_Mesh'. SelfPackage differs because the mirror lives under /Game/Decompiled/…; normalization produces different strings for the same absolute reference.
2. EditorData / EditorParameters: names such as …破空灰尘:Fountain.NiagaraEmitterEditorData~ versus …破空灰尘:Fountain~.NiagaraEmitterEditorData~ reflect editor-only object naming.
3. bBindingExistsOnSource=True→False inside renderer *Binding fields is compiler-derived cache noise. Do not extend this dismissal to MaterialParamValidMask: the section 3.6 difference needs independent investigation.

3.11 Explicitly unverified features

Stage blocks; more than one event handler or event generators; custom C++ stage classes; RibbonRenderer; SubImageSize; custom pivot/scale/LOD in MeshRenderer.Meshes elements; EffectType; ModulePaths; RequiresPersistentIDs; as <node name>; Defaults = {}; actual inline hlsl { }; .dfe from references; and per-device-profile platform-set overrides (FNiagaraPlatformSet beyond QualityLevelMask). Absence from this content is not proof of support.

4. Supported write-back inventory

A feature counts as measured write-back support only when it appears in the export, passes L1/L2, and survives reflection-level facts. The first two alone establish only text-level success, with reflection coverage still qualified below.

5. Fixed-point samples

Five systems were sampled because bulk rebuilding can stall the editor: the documented example has 24 systems creating 237 compile jobs. Each source was rebuilt with -Force in a separate editor process.

The aggregate scan reported L1: 5 passed, 0 failed, 54 never built; L2: 5 compiled clean, 0 failed. dfx returned FAILED (exit 54) because Missing contributes to the exit code (3.9). asset-diff reported 2 same, 3 different, 54 missing. These aggregate counts cover the available mirrors in the scan and are distinct from the five newly selected build attempts.

The central result is that mirror-diff passes all five available mirrors while asset-diff finds differences in three assets.

Additional S2 evidence: re-exporting its mirror and using Compare-Object against the original export gives 103 lines versus 103, differing only in the first-line provenance comment. Even MaterialParameters, Meshes and OverrideMaterials blobs match byte for byte. Its problem is not the text round trip, but the text's omission of MeshYaw.

6. Implementation recommendations and follow-up

The separate DreamGUI research note proposed this order: (1) build safety gate, (2) pull, (3) line-level write-back plus dirty tracking, (4) coordinated compile suppression. This section adds measured evidence and specific changes, revising the basis for the first two items.

6.1 Safety gate: use DescribeSystemFacts, not exported text

The earlier suggestion was to reuse the DFX8010 gate in build. The measurements support the goal but not that implementation basis: DFX8010 relies on Result.UnsupportedFeatures, yet the three differing assets reported none of the losses in 3.4–3.7. S2 and S5 report zero; S3 reports only the four declared parent links. A text-level gate cannot see these losses.

DescribeSystemFacts in Source/DreamFXEditor/Private/Diff/DreamFXAssetFacts.cpp is already an independent reflection-based asset probe. After loading the target and before committing writes:

1. Record its current facts.
2. Determine what the source will produce.
3. List facts that the text cannot reproduce and refuse destructive replacement unless explicitly forced with -Force.

This reuses an existing tool at one integration point to cover the S2/S3/S5 classes of silent loss.

6.2 Add system properties to asset-diff

DescribeSystem (DreamFXAssetFacts.cpp:246) never calls AppendPropertyFacts for the UNiagaraSystem object itself. Reflecting that object adds an independent witness for EffectType, WarmupTime, FixedBounds, bFixedTickDelta and bDeterminism. Removing VersionedParent (:258) from the skip set was also proposed so parent-chain comparison would not rely on a separate adapter path.

6.3 Fix <Emitter>.<Node>.<Input> read/write asymmetry

IsNamespacedName (DreamFXValueLowering.cpp:204-227, allowlist at :39-44) rejects references whose first segment is an emitter name, although the decompiler emits them (:655-656). Options:

• Accept real context: Bind already accepts emitter aliases (DFX4026 comment in DreamFXGenerator.cpp:1490-1502), and Output.<node>.<value> is already allowed. Recognizing emitter/node names from this system is more accurate than expanding a fixed keyword list.
• Declare unsupported output: if it cannot be expressed, record UnsupportedFeatures instead of emitting a writer-rejected token, so coverage stops describing it as ok.

Either approach must make the read-coverage report acknowledge current rebuild limitations.

6.4 Read path: export every input explicitly

pull and dirty tracking require an input's current value even when it equals its default. Initially only -NoDefaults did this, while its documented contract excluded maintained source (.skill/dfx.ps1:127-129). The implementation already exists as FDecompileOptions::bIncludeDefaultedInputs (DreamFXCommandlet.cpp:265, 1100).

The initial recommendation was to promote this to the read mode for write-back, and define the difference between normal and -NoDefaults exports as a suppressed set. Dirty write-back would leave that set alone and update only keys the user actually changed. Otherwise the first flush could erase values described in 3.4/3.5. Section 6.9 records the more direct implemented approach.

6.5 Investigate MeshYaw first

This is the one definite source/asset value mismatch with related evidence suggesting a visual effect. First compare original and /Game/Decompiled/…/NS_Effects1_Mesh visually or through L3. If rendering differs, add a read-path assertion comparing the input's reported value with its rapid-iteration stored value, and report a gap when they disagree. Pin/store conflicts merit their own diagnostic rather than silent export.

6.6 Correct gate interpretation

• State explicitly that coverage measures read coverage, or add a rebuildability column. One of these 59 systems reads successfully but cannot rebuild.
• mirror-diff includes Missing in its exit code. Either distinguish missing mirrors with a dedicated result, build the whole tree in CI, or provide an “already built mirrors only” mode.
• Add asset-diff to the fixed-point gate. Byte-identical no-op write-back alone is insufficient: three measured assets have reproducible text but unequal asset state. The initial acceptance proposal was both byte-identical L1 and equal asset fact sets; section 6.10 refines that criterion.

6.7 Documentation corrections suggested by measurements

6.8 Implemented: build safety gate

Branch feat/build-safety-gate, diagnostic DFX8018. It implements section 6.1 with DescribeSystemFacts and adds the system-property coverage from 6.2. These are follow-up measurements; the initial measurements above remain historical evidence.

Second-pass semantics: detect drift in state the source does not control.

The first version (66db8ce) blocked every fact not retained exactly after rebuilding. That immediately blocked ordinary authoring: changing a declared UniformScale from 1.0 to 2.0 looked like loss of 0000803F, and deleting a module was blocked. The second version (b3ca74b) compares each structure and consults source declarations first:

• Before saving: collect facts before the first asset mutation, then again after compilation and provenance stamping but before SaveSystem. Use reflection-level multiset difference, not just the first mismatch.
• Rule 1 — compare only structures present on both sides. An emitter, module node or renderer present only on one side is an intentional source addition/deletion and is ignored. Deleting an entire module therefore succeeds.
• Rule 2 — a missing fact inside a retained structure is blocked. The source could not represent that state; saving would destroy it, as with a lost material binding. One later exception (6.10): if the same value remains at the same address after removing its script label, the missing instance was a duplicate in another script store. Allow the collapse and report verbose merged |; names alone are insufficient.
• Rule 3 — a changed fact inside a retained structure is blocked only when undeclared. Consult parsed declarations, not guessed names. Settings use top-level keys from the planned JSON, exactly what the generator writes. Module inputs use AST argument names and Niagara variable names registered in ApplyStack against the actual engine-assigned node name. Only there is the name after as <name> known; guessing “module asset name plus counter” is explicitly rejected.
• Rule 4 — uncertainty blocks, with distinct wording. Details end in [deterministic drift (the text does not name this input)] when declarations were inspected, or [suspected drift (cannot tell)] when no declaration record exists for that structure.
• Self-contained fact families: di, compiled, user, event handlers, Stage and emitters without data lack registerable source addresses. Each whole fact is treated as its structure, so rule 1 interprets differences as whole-record additions/deletions. This is an explicit limit: DI/compiled facts lack node names, and Stage internal properties do not map one-to-one to Stage(...) argument spellings.
• -Force still saves, emitting the same details as warnings with the same markers.
• A new target asset is not gated. Neither are verify and -NoSave, which do not save.
• Compare stable facts: strip bBindingExistsOnSource, bIsCachedParticleValue and empty compiled … writes: lists caused by PostLoad discarding cached VM data. These changes affect only the gate; asset-diff output and the earlier recorded counts remain unchanged.
• New collection surface: system-object properties now appear as system WarmupTime = …. The skip list excludes compiler products, editor bookkeeping, identity and R3-extracted scratch pads.
• Only explicit -Force bypasses the gate, independently of ordinary bForce provenance bypass. Adopt, watcher rebuilds and the DreamGUI bridge set the latter only. Normal declared edits pass automatically; genuine loss is blocked.
• Artifacts on every trigger: Saved/DreamFX/BuildSafety/<asset>.{before,after,lost}.facts. After forcing, the asset no longer contains the previous state, so these files preserve the evidence.

Measurements used the same compiled plugin, one editor process per command. Logs: Saved/DreamFX/gate-probe/runs5|runs6/.

Count correction: section 3.4's “37+4 original-only facts” compares an original with its mirror. Rebuilding in place, cleaning stale parameters then refilling from source, loses only the eight facts in row 7. The other 29 module constants equal defaults and are restored as identical facts. “The source omits them” is true; “an in-place rebuild loses all 37” is not.

A deliberate boundary authors may disagree with: removing an argument while keeping its module still blocks under rule 3, because the source no longer specifies that input (row 4). Deleting the entire module passes because the source explicitly removes the whole structure. If “reset this field to its module default” should also count as intent, the relaxation belongs in the single Declared.Names decision.

Row 8 exposed a useful secondary finding: maintained template source DFX/Templates/NS_Atlas2D_Mesh.dfs lacks the MaterialParameters = "{…}" blob present in the mirror export. Rebuilding it using the documented template workflow clears the Sheet binding and recreates the white-block problem. The gate catches it; that workflow now requires explicit -Force, followed by the AtlasFXSetup step already requested in the template header. ci.ps1 uses build -All -Force, so CI still proceeds with loss reported as warnings.

Associated fix required by the bundled corpus: collecting system facts exposed RoundTrip/ValueModes.dfs, which declares FixedBounds = box(-50…50) but rebuilt as default ±100. The writer set the box without enabling bFixedBounds, while the reader emitted it only when enabled. Other fixtures all used default ±100 and had hidden this. A companion bool on the system FixedBounds mapping (FSettingMapping::CompanionBool) fixed it. Consequence: the only project sources declaring system bounds, DFX/Templates/NS_Atlas2D_Mesh.dfs and NS_Atlas2D_Sprite.dfs (±300), will actually enable fixed bounds on their next rebuild. This matches their declared intent but is a behavior change. The repaired corpus passed 55/55.

Write audit: only two disposable probes were created, /Game/DreamFXProbe/NS_GateProbe and /Game/DreamFXProbe/NS_GateSysProbe (removed after validation), plus rebuilt mirrors. The three targeted originals, NS_Effects1_Mesh, 破空灰尘 and /AtlasFX/Templates/NS_Atlas2D_Mesh, retained their hashes because refusal happened before save. Byte backups are in Saved/DreamFX/gate-probe/asset-backup/.

6.9 Implemented: pull (asset → source)

Branch feat/pull; dfx pull; diagnostics DFX7106–DFX7116. The pull reference contains full semantics, limits and measurements. This section relates them to the earlier findings.

• Section 6.4 was partly adopted and partly replaced. -NoDefaults still puts keys in the source; pull cannot update an undeclared input. Values now come directly from reflection over the rapid-iteration store, aligned with asset-diff and the gate, rather than from pins. No separately computed suppressed set is needed: undeclared source inputs are simply left alone (DFX7106).
• Section 3.5's MeshYaw can now be corrected in source. pull -Apply changes exactly one line, 0.0 -> -90.0; build then reports one built, zero errors and no build safety messages, since rule 3 allows the declared value. Both asset-diff sides become 0000B4C2. The L3/visual question remains unverified: this fixes whether the source contains the value, not which runtime representation is consumed.
• The eight facts from 6.8 row 7 are reported and skipped, demonstrating that ordinary pull does not add structure. For 破空灰尘: 19 compared, zero to write, 94 undeclared, zero unaddressed stacks, unchanged SHA256. Counting distinction: 94 is deduplicated by name; the earlier 37/57 numbers count per-script facts, including separate spawn/update copies of one name. They are different quantities and must not be interchanged.
• Adapter contract discovered by pull: FNiagaraAdapter::GetEmitterInfo appends to its output. Reusing one FEmitterInfo for two emitters leaks the first emitter's stacks into the second. pull creates a fresh instance per emitter and documents this. The shared adapter itself was not changed; that requires a separate assessment.

6.10 Implemented: wider coverage, dirty tracking and structural edits

Branch feat/write-back, three commits starting from feat/pull; new diagnostics DFX7112–DFX7115. Details and measured cases are in pull, sections 4b/4c/4d/5b.

Disposition of the earlier gaps:

• 3.4, omitted module literals: decompile -NoDefaults still supplies input keys; ordinary pull adds no arguments. The formal value path directly reads the rapid-iteration store for source-declared addresses. The separate suppressed-set proposal remains unnecessary.
• 3.5, MeshYaw: already addressed for text write-back in 6.9; unchanged this pass.
• 3.6/3.7/3.8, MaterialParamValidMask, DI node names and EffectType: untouched this pass. These concern renderer/compile state, node identity and a missing reader field, respectively, rather than a value write-back operation.

The initial support table understated later coverage. Renderer properties, settings and assignments now have assertion-backed write paths. Settings uses the generator's own mapping table in reverse; renderer value spelling is shared by decompiler and pull through WriteBack/DreamFXSourceValue.h. Extracting that shared implementation exposed a regression: leftover break statements from the previous switch caused only the first property to export, failing five RoundTrip fixtures. Independent DescribeSystemFacts assertions caught what a symmetrical export comparison could miss.

The asset-level acceptance criterion was implemented more precisely: after rebuilding the same asset in place, no fact may disappear while its address remains. This targets the same state-loss category as DFX8018 rather than requiring two separately created assets to have identical facts. Removed whole structures and collapsed duplicate constants are reported separately. The default-value restoration described in 6.8 explains why missing source entries do not all become lost facts on an in-place rebuild.

A measured difference in criteria, resolved by feat/collapsed-copy-drift. SetInput, the same external-edit API used by editor sliders, writes an EmitterUpdate constant into two stores, system-spawn and system-update, while a DreamFX rebuild materializes one. The previous pass recorded this as an unresolved discrepancy. Probe artifacts and complete logs are in Saved/DreamFX/CollapseProbe/; the real-system export was /AtlasFX/Decompiled/Templates/NS_Atlas2D_Mesh, editing EmitterState's Loop Duration.

1. It did block, with the predicted wording. After a slider edit, pulling in the same session then rebuilding produced:

   build safety: '…' held 187 fact(s), the rebuild produces 195; of the 1 fact(s) it no longer holds exactly, 1 are drift inside a structure the rebuild kept, 9 gained.
   lost | ri system-spawn Constants.Atlas2D_Mesh.EmitterState.Loop Duration (NiagaraFloat) : 0000A040 -> one copy fewer
   error DFX8018: … was not saved …

   Raw log: CollapseProbe/A3-refusal-before-fix.log.

2. The trigger window is narrower than initially suspected. The duplicate lives only in the session that wrote it. Even the same session loses its ri system-spawn copy after SaveSystem: measured 187 facts / 0 system-spawn before editing, 188 / 1 afterward, 189 / 0 after save. A fresh process also reads only one. Thus slider → Ctrl+S → rebuild does not trigger it (CLI A1/A2 builds passed). The failing path is edit a value in the editor, make text match the store in that same session (pull or manual edit), then rebuild without saving the asset. That is precisely the bidirectional workflow under discussion.

3. Editor entry points do not bypass the gate. FGenerateOptions::bForce skips provenance; bForceLossyRebuild allows loss. Watcher, DreamGUI bridge and Adopt set only the former, as explained in DreamFXGenerator.h. A .dfs save triggering a watcher rebuild in that session can therefore be refused, with no CLI -Force switch at that entry point.

4. Fix: a rule-2 exception keyed by value plus script-independent address. FBuildSafetyGate::Compare checks which (address,value) pairs survive. The address is Constants.<emitter>.<node>.<input>; the script label only says where the value is stored. For the measured EmitterUpdate constant, editing adds a same-byte system-spawn copy of a system-update value, and rebuilding retains one. Same address and value still present means a representational duplicate collapsed, not state loss. Allow it but emit verbose merged | … and the summary a further N cop(y/ies) of a constant another script's store still carries collapsed into one -- not drift. Only equal values qualify. Different values or a truly absent value still use rules 2/3/4, with unchanged (missing), deterministic and suspected markers. The sole relaxation is the AfterValuesByKey lookup, not a blanket skip of names containing system-spawn.

5. Corpus assertions: new DreamFX.Corpus.WriteBack.CollapsedCopy tests four fact pairs directly against Compare, asserting decisions rather than reimplementing them: two equal copies become one → zero blocked, one collapsed; unequal copies with the old value gone → one deterministic block; the same unequal case with the input declared → zero blocked under unchanged rule 3; retained address but entire subject absent → one (missing) block. ModuleInput now asserts, rather than merely reports, 0 lost / 1 collapsed on rebuilding:

   ModuleInput: rebuilt in place … 172 fact(s) before, 171 after, 0 lost, 1 duplicate copy/copies collapsed, 0 structure(s) the text removed

6. The symmetric pull case needs no semantic change. FindStored in WriteBack/DreamFXPull.cpp deduplicates by name: equal copies become one source value; unequal copies are refused because pull cannot determine which the engine uses (pull section 6). Equal values therefore yield one source literal, one rebuilt store value, then an idempotent pull (DFX7107 … already holds every value … No bytes written, unchanged SHA256 and mtime). Unequal values are not pulled; build writes the source's intended value to its own store. The operations do not repeatedly overwrite each other's choices. Pull semantics did not change.

End-to-end measurements: the disposable /Game/DreamFXProbe/NS_PullProbe was removed after verification. The seven cases are in pull section 5b. The structural rebuild succeeded without any safety-gate diagnostics and did not create Saved/DreamFX/BuildSafety/; its only warning in the detailed log was pre-existing DFX7102 for unbounded SpawnRate. No-op pull -Apply preserved both SHA256 and mtime. When text changed and the asset did not, DFX7115 preserved the text's newer value without touching the file.

7. Reproduction

Environment

• Project C:\CrossingVoid; engine D:\UnrealEngine-5.8.2 with the DreamFX engine patch.
• Plugin C:\CrossingVoid\Plugins\DreamFX at 359b3b3, branch pr/engine-patch-and-writeback.
• Ensure no UnrealEditor is running before reproducing. None was running at the start of this study.

Commands, in execution order

$dfx = 'C:\CrossingVoid\Plugins\DreamFX\.skill\dfx.ps1'
$eng = 'D:\UnrealEngine-5.8.2'

# 1) Coverage: 59 systems, 59 exports, 0 failures, 17 gaps.
pwsh -File $dfx coverage -Path '/AtlasFX+/Game/GameActor2D+/Game/AssetMaterial' -Engine $eng

# 2) Full export: 59 written, 0 failed, to C:\CrossingVoid\DFX\Decompiled\**.
pwsh -File $dfx decompile-all -Path '/AtlasFX+/Game/GameActor2D+/Game/AssetMaterial' -Engine $eng

# 3) Fixed-point sample: one build per source; write only Decompiled/ mirrors, never originals.
#    See Saved\dfx-writeback-study\run-builds.ps1.
pwsh -File C:\CrossingVoid\Saved\dfx-writeback-study\run-builds.ps1
#    S1 NS_Atlas2D_Mesh OK; S2 NS_Effects1_Mesh OK; S3 TR刀光_System OK;
#    S4 NS_ScanScene failed with 21 errors; S5 破空灰尘 OK.

# 4) L1/L2 fixed-point check.
pwsh -File $dfx mirror-diff -Path '/AtlasFX+/Game/GameActor2D+/Game/AssetMaterial' -Engine $eng

# 5) Independent reflection comparison, persisting facts.
pwsh -File $dfx asset-diff -Path '/AtlasFX+/Game/GameActor2D+/Game/AssetMaterial' -DumpFacts -Engine $eng

# 6) Print suppressed inputs, the key step in section 3.5.
pwsh -File $dfx decompile /Game/GameActor2D/Misaka/Material/Sk1/NS_Effects1_Mesh -NoDefaults -Out <out>.dfs -Engine $eng

Artifact locations

Change audit for the original study

• No original assets changed. Builds wrote only mirrors in Decompiled/. The only extra non-text export side effect was R3 embedded-script extraction to /Game/Decompiled/…/Scripts/NS_ScanScene_*.uasset, also under Decompiled/.
• The project already had modified Plugins/AtlasFX/Content/Templates/NS_Atlas2D_*.uasset files (mtime 17:47) and Content/GameActor2D/Misaka/Material/DefAtk/FX_Defatk.uasset (mtime 18:07), preceding the study's 20:20 start. These were pre-existing changes.
• The plugin's git status was clean before this report was written.
• D:\UnrealEngine-5.8.2 was unchanged.