fix: memory leak in composite drag and drop registrations

[Simon Siefke (SimonSiefke)]

Details

Each drag source or drop target registration adds its disposal store to the process-lifetime observer as well as returning it to the caller. Disposing the registration clears its callbacks but leaves the empty store and its Set in the observer.

Change

Return each registration store to its existing caller owner without adding it to the observer’s lifetime store. The caller continues to dispose its drag source or drop target normally.

Before

Toggling Modern UI on and off 37 times after warmup grows the observer’s disposed registration stores from 471 to 915. The full aggregate Set measurement below grows from 26,550 to 26,987 entries; it includes unrelated Sets too.

After

The same 37-cycle test leaves no registration stores in the observer. The full aggregate Set measurement stays at 26,051 entries, and overall DisposableStore growth falls from 447 to 4. Undo history and other object-shape differences remain. This fixes empty bookkeeping objects, not callbacks or widgets that disposal already released.

Test Video

Seven Modern UI enable/disable cycles in Settings after the same setup warmup.

test.mp4

_{AI disclosure: Model: GPT 6 Astra. Worktime: 51 min}

[Copilot]

Copilot review overview

🟢 Approval recommended

All production callers retain and dispose the returned registrations, and the tests cover lifecycle regressions.

Review effort: Balanced
Findings: None

What changed in this PR

Prevents process-lifetime retention of disposed composite drag-and-drop registration stores.

Changes:

• Returns registration stores directly to caller owners.
• Adds GC, disposal, and drag lifecycle regression tests.

File	Description
src/​vs/​workbench/​browser/​dnd.ts	Removes observer ownership of per-registration stores.
src/​vs/​workbench/​test/​browser/​compositeDragAndDropObserver.test.ts	Tests collection, disposal, and event behavior.

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Dmitriy Vasyura (dmitrivMS)]

Simon Siefke (@SimonSiefke) Thank you!