[bobhenl]

• PlayerUI skulls of players without skin textures no longer get a profile the server has to look up. A textureless skull profile makes the server run PlayerProfile.update() on every build (and GuiItem re-applies the meta on every render). On offline-mode servers those Mojang lookups for v3 UUIDs never succeed, so they just pile up.
• That plain-head branch is the effective change: untextured non-v4 players get a plain head, setOwningPlayer stays only for v4 UUIDs. Textured players get their live PlayerProfile, which on real Paper is behaviour-neutral (setOwningPlayer(Player) already uses the live profile).
• The player name is stored in the head's PDC, so plain heads stay selectable and filter formats still apply to them.
• The cache keeps the head and the skin/cape it was built from in one entry. A head is rebuilt when the textures change (skin arriving late from a skin plugin, or a skin change mid-session).
• Offline players and failed builds leave no cache entry.
• Tests in HeadCacheTest and PlayerUIScreenTest, including v3 untextured players. The first commit has run on our network (3 Paper/Leaf 26.3 servers, offline mode behind Velocity) since Oct 29 without issues.

[CyR1en]

Thanks for this. I checked the diagnosis against the Paper 26.1.2–26.3 sources and it holds: CraftMetaSkull.applyToItem calls update() whenever the owner profile has no skin/cape, that goes straight to an uncached fetchProfile(uuid, true), and Mojang answers 204 for v3 UUIDs. It is actually worse than described, because GuiItem.applyUUID() re-applies the meta on every GUI render.

Requesting changes for one regression:

Plain heads can't be selected. PlayerUIScreen identifies the clicked player with skullMeta.getOwningPlayer() (PlayerUIScreen.java:132-134). The new plain head has no owner, so clicking an untextured non-v4 player does nothing. applyFirstFilterFormat (lines 261-263) has the same dependency, so filter formats are skipped for those heads too. Both reproduce in MockBukkit with a v3 player and pass on development. CI misses it because the default mock players get random v4 UUIDs.

Suggested fix: write the player name (or UUID) to the head's PDC in createHead and read that in both places, with tests using a v3 untextured player.

Smaller points:

• hasTextures checks for a non-empty property, but the server triggers on "no skin and no cape URL". !profile.getTextures().isEmpty() mirrors it exactly.
• cache and untextured are updated in two steps (invalidate, clear), which can drop a marker under Folia. A single map with a (item, textured) value would remove the invariant, or use synchronized (cache).
• untextured.add happens before the build completes, so an exception leaves a marker with no entry.
• On real Paper, setOwningPlayer(Player) already uses the live profile, so the textured branch is behaviour-neutral. Worth saying in the description that the effective change is the plain-head branch.
• Nits: createHead returns null from an Optional method; the map can no longer hold Optional.empty(); the Javadoc says "version 3" but the check is "not version 4".

Optional follow-up: compare the cached textures value with the live one so mid-session skin changes are picked up as well.

[bobhenl]

Thanks for the thorough review. Pushed 7a9d8e5:

• Plain heads are selectable again: createHead writes the player name to the head's PDC (commandprompter:head_owner), and both the click handler (now selectHead) and applyFirstFilterFormat read it through HeadCache.getOwnerName, falling back to getOwningPlayer. New tests with a v3 untextured player cover the click and the filter format; with the PDC write removed they fail.
• hasTextures is gone; the check is now !profile.getTextures().isEmpty(), same as the server.
• untextured is removed. The cache is a single map of (item, textures), and the compare-and-rebuild in getHeadFor runs under synchronized (cache). Nothing is stored if the build throws (test added).
• Follow-up included: the cached textures key (skin + cape URL) is compared with the live one, so a mid-session skin change rebuilds the head.
• Nits: createHead returns a nullable record instead of a null Optional, the map no longer holds Optional at all (the empty-entry tests are replaced), and the Javadoc now says non-version-4.

I also updated the description: on real Paper the textured branch is behaviour-neutral, the effective change is the plain-head branch.