fix(textures): evict orphaned cached textures during cleanup

[chiefcll]

What

Adds an orphan-eviction sweep to TextureMemoryManager.cleanup() so keyCache entries are bounded again, plus a new EventEmitter.hasListeners() liveness signal and unit tests.

Why

Since #87, cleanup() only ever calls the reversible freeTexture() (so textures reload in place), which left destroyTexture() with zero callers — nothing evicts keyCache entries anymore. GPU/CPU image data is still released on free, but each retained Texture JS object is a slow leak in long-running TV apps cycling many unique image URLs (infinite catalog rows).

How

evictOrphanedTextures() runs at the end of every cleanup() and destroy-and-evicts a cached texture only when nothing can ever re-display it:

• zero renderableOwners, AND
• zero event listeners (every live referencer — CoreNode.loadTextureTask, SubTexture — subscribes via on() and unsubscribes on teardown), AND
• not preventCleanup, AND
• in an evictable state:
  • 'freed' — already reclaimed by a prior cleanup, or
  • 'initial' / 'failed' — only after the existing 2s startup grace period, which covers the same-frame race where a brand-new node's listener subscription is still in the microtask queue. Nothing retries a 'failed' texture without a failed listener, so those are pure dead weight.

In-flight states ('fetching'/'loading'/'fetched') are never evicted, so slow network downloads are unaffected (they also hold listeners and usually an owner the entire time). 'loaded' orphans go through the pressure-driven free path first and are swept as 'freed' on a later pass.

Invariant preserved (the #87 blank-poster bug)

A texture with any listener attached is never destroyed here. destroy() calls removeAllListeners(), which would sever a live node's subscription — the exact bug #87 fixed. Nodes only subscribe once in loadTextureTask, so the skip direction is conservative: when referenced, keep.

One benign residual race: a new node can cache-hit an old orphan in the same frame as a cleanup (grace period doesn't protect stale createdAt). Worst case is a lost cache entry → a possible duplicate texture instance; the node subscribes after the destroy, so its listeners survive and the texture still reloads and renders.

Testing

• 18 new unit tests in src/core/TextureMemoryManager.test.ts: referenced-freed kept (listeners / owners / preventCleanup), orphaned-freed evicted, full lifecycle (kept while subscribed → evicted after off()), same-pass free→evict, fresh 'initial' kept during grace, aged 'initial'/'failed' orphans evicted, aged 'initial' with listeners kept, loaded/in-flight states never evicted, and hasListeners() across on/off/once/removeAllListeners. Eviction fakes are built on a real EventEmitter so listener accounting matches CoreNode exactly.
• vitest run: 396 tests pass. pnpm build, prettier, eslint clean.
• No rendering change, so no new visual-regression test; the existing suite covers the reload-in-place path.

🤖 Generated with Claude Code