New Scene destroys the reused UI FontAsset's atlas (2026-08-03)

Live report: “NewSceneを作成したらまたエラーが発生してしまってます” — the v0.13.1/v0.14.0 font work stopped the warning flood, but creating a new scene produced a fresh flood of MissingReferenceException / ArgumentNullException / NullReferenceException out of TextCore during every UI Toolkit repaint of the panel.

1. Symptom (live editor, DevelopmentProject)

uloop get-logs --log-type Error returned 12 distinct errors, all rooted in UIRStylePainter.DrawText → UITKTextHandle.Update → TextCore:

#	error
0,1	MissingReferenceException: The variable m_AtlasTextures of FontAsset doesn't exist anymore (in FontAsset.TryAddCharacterInternal)
2-5,7,8	ArgumentNullException: Value cannot be null (in MaterialManager.GetFallbackMaterial)
6	MissingReferenceException: The variable m_Material of FontAsset doesn't exist anymore
9,11	MissingReferenceException: The object of type 'Material' has been destroyed (in MaterialManager.GetFallbackMaterial)
10	NullReferenceException in UIRStylePainter.DrawTextInfo

2. Root cause, measured

2.1 The live asset’s state

A read-only execute-dynamic-code probe against the running editor:

== 'UAP_JA_UI_FONT' hideFlags=HideAndDontSave popMode=DynamicOS persistent=False
   material:  refnull=False unityNull=True      <- dangling: destroyed
   atlasTextureCount=4 multiAtlas=True
   tex[0..3]: refnull=False unityNull=True      <- dangling: destroyed
FRESH CreateFontAsset: hideFlags=None / material hideFlags=None / tex[0] hideFlags=None

So the FontAsset itself survived (it carries HideAndDontSave, which includes DontUnloadUnusedAsset), while every child object it owns was destroyed. FontAsset.CreateFontAsset hands back children with hideFlags = None; v0.14.0 stamped the parent only.

2.2 What destroys them — isolated in the sandbox

Sandbox AITemp, batch mode, Assets/Editor/UapFontProbe.cs. Asset A = v0.14.0 behaviour (parent stamped only); asset B = parent and children stamped. Both grown to 9 atlas textures via TryAddCharacters:

== step 1: EditorUtility.UnloadUnusedAssetsImmediate() ==
    A  mat=alive  all used textures alive
    B  mat=alive  all used textures alive       <- NOT the culprit
== step 2: EditorSceneManager.NewScene(DefaultGameObjects, Single) ==
    A  mat=DEAD   9/9 used textures DEAD        <- the culprit
    B  mat=alive  0/9 used textures DEAD        <- stamping is a complete fix
== step 3: second NewScene ==
    A  mat=DEAD   9/9 DEAD
    B  mat=alive  0/9 DEAD
== step 4: reuse the broken A ==
    A.TryAddCharacters(...) returned False      <- unusable, not recoverable

Three findings that shape the fix:

1. Resources.UnloadUnusedAssets is innocent. Only the scene unload performed by NewScene destroys the children. Guessing at “UnloadUnusedAssets probably did it” would have produced the wrong fix.
2. HideAndDontSave on the children is sufficient and complete.
3. A broken asset cannot heal. TryAddCharacters returns false, and because the FontAsset carries HideAndDontSave it also survives every subsequent domain reload — so once broken it stays broken for the whole editor session, and FindExistingMarkedAsset keeps handing it back.

2.3 Stamping once at creation is NOT enough

Same probe, asset B was stamped immediately after creation and then grown:

B atlasCount=9  texFlags=[HideAndDontSave, None, None, None, None, None, None, None, None]

TextCore’s multi-atlas growth creates each new atlas texture with hideFlags = None. Atlas textures appear lazily, during text generation, as new glyphs are rasterised — there is no creation hook to piggyback on. The fix therefore needs a cheap recurring re-stamp, not a one-shot.

2.4 The atlasTextures array is over-allocated

atlasTextures.Length was 16 while atlasTextureCount was 9; the trailing 7 entries are legitimately null. A health check must iterate atlasTextureCount, not atlasTextures.Length, or every healthy asset reads as broken. (The first draft of the probe made exactly this mistake.)

2.5 TextCore’s fallback materials are NOT a second error source

Errors 2-5 and 7-11 come out of MaterialManager.GetFallbackMaterial, so the obvious worry was that its static cache of per-atlas materials is destroyed too — which we could not fix by stamping our own asset. Probe #3 drove that cache directly through reflection:

fallback[1..4] = hideFlags=HideAndDontSave (created that way by TextCore)
cache BEFORE NewScene: count=4 alive=4 destroyed=0
cache AFTER  NewScene: count=4 alive=4 destroyed=0
re-request fallback[1] = alive

TextCore already protects them. The GetFallbackMaterial failures were about the sourceMaterial argument — our FontAsset’s own destroyed material — so fixing 2.2 fixes all 12 errors, not just the first two.

3. Options considered

#	option	verdict
1	Go back to destroy-and-recreate on beforeAssemblyReload	Rejected. That is precisely the v0.13.x defect (1106× MissingReferenceException) that v0.14.0 fixed, and it would not help here at all: NewScene is not a domain reload.
2	Disable multi-atlas (isMultiAtlasTexturesEnabled = false) so only one, creation-time-stamped texture ever exists	Rejected on measurement. Probe #2: a single 1024×1024 SDFAA atlas at the default 90pt holds 104 CJK glyphs (2896 of 3000 requested went missing). Unusable for Japanese UI.
3	Re-stamp only from scene-lifecycle callbacks (sceneClosing, newSceneCreated, …)	Rejected. Requires enumerating every code path that can unload a scene; a missed one silently reintroduces the bug.
4	Stamp children at creation/reuse + cheap recurring guard + reject unhealthy assets at resolve	Chosen.

4. Chosen design

1. StampSubObjects(FontAsset) — sets HideFlags.HideAndDontSave on the asset’s material and on every non-null entry of atlasTextures. Idempotent; only writes when the flag differs.
2. Stamp at both entry points — right after CreateFontAsset, and on the re-find-and-reuse path (an asset created by an older version arrives with unstamped children).
3. Recurring guard — an EditorApplication.update hook, installed lazily the first time an asset resolves, that re-stamps when atlasTextureCount has changed since the last stamp. The guard is an int comparison per editor frame in the common case. A count change is the only way a new atlas texture can appear, and it also catches ClearFontAssetData (which resets the count to 1).
4. IsHealthy(FontAsset) — material alive and the first atlasTextureCount atlas textures alive (see 2.4).
5. FindExistingMarkedAsset rejects unhealthy assets. A marked asset whose children were destroyed by a pre-fix session is provably unusable (2.2 step 4) and would otherwise persist for the whole editor session, so it is destroyed at the same point the dedupe path already destroys extras, and a fresh asset is created. This is safe for the same reason dedupe is: resolve only runs on the first request of a domain load, before any live VisualElement has been styled with the result.

Users already running a broken editor therefore recover on the domain reload that installs this version — no editor restart required.

5. Regression guards

Tests/Editor/FontLoaderTests.cs:

• StampSubObjects_MarksMaterialAndEveryAtlasTexture — the property that 2.2 proves implies survival, plus idempotency (the guard runs per frame).
• GrownAtlas_NewTexturesComeBackUnstamped_AndSpareSlotsStayHealthy — one growth pass pinning both 2.3 and 2.4 (rasterising is the slow part, so the two assertions share it).
• IsHealthy_FalseWhenMaterialDestroyed / ...WhenUsedAtlasTextureDestroyed / IsHealthy_FalseForNull.
• FindExistingMarkedAsset_RejectsAndDestroysUnhealthyAsset — pins 4.5.
• AtlasGuardTick_ReStampsOnlyWhenAtlasCountChanged — pins the guard’s cheap-path contract in both directions.
• JapaneseUiFontAsset_ResolvedAssetHasStampedSubObjects — end-to-end.

The NewScene causation itself is not in the suite: a scene-unloading test would disturb the other 1450+ EditMode tests. It is pinned by the sandbox probe transcript reproduced in section 2.2 instead.

6. Verification

• Sandbox AITemp: compile clean, full EditMode suite 1466 total / 1461 passed / 0 failed / 5 skipped (was 1458 before the 8 new tests; all 8 ran, none ignored).
• Live editor (DevelopmentProject), after uloop compile:
  • The broken asset was rejected and destroyed, a fresh one created (JapaneseUiFontSource = "osasset:Yu Gothic UI", i.e. the create path, not the reuse path), material and tex[0] both HideAndDontSave, exactly one marked asset alive. Recovery needs no editor restart.
  • Recurring guard, end to end: forcing 200 CJK glyphs grew the atlas to 3 textures and tex[1] tex[2] read None in that same frame; a later frame read HideAndDontSave on all three. The spare slot (arrayLen=4, atlasTextureCount=3) stayed null and the asset stayed healthy, exercising 2.4 in production.
  • uloop get-logs --log-type Error went from 12 errors to 0.