VS Code extension: a single prompt submission was dispatched into 3 separate sessions (duplicated token usage)

**Environment**
- Claude Code VS Code extension (`entrypoint: "claude-vscode"`)
- Windows 11

**What happened**

A single message I sent from the VS Code extension was processed as three independent sessions running in parallel, each executing the full task from scratch. All three sessions consumed a full agentic run (hundreds of assistant turns each), roughly tripling token/quota usage for what was intended as one request.

**Evidence (from local `~/.claude/projects//*.jsonl` transcripts)**

Three session files were created within the same second:
- Session A: `3cb5e67c-9a1d-4ccd-abca-81e1efb97fcf`
- Session B: `99a4d3be-9f1c-471d-a042-ae7578c82994`
- Session C: `454c035f-6482-4afa-a1ee-d6d120e4f092`

Key correlating identifiers that are identical across sessions A and C (indicating they originated from the *same* user submission, not two separate messages I typed):
- `promptId`: `127520aa-e797-4ee8-b68b-7502f86c1e3d` — identical in both session A and session C's first user message
- User-message `timestamp`: `2026-09-15T09:02:22.868Z` — identical to the millisecond in both
- User-message content length: identical (48702 chars) in both
- `cwd` identical in both
- SessionStart hook `toolUseID`: `0ec5dd30-2a2b-4e60-8f23-3593939dc83d` — identical across **all three** sessions' first log line

Session B additionally logged an explicit queue race right at the same moment:
```
{"type":"queue-operation","operation":"enqueue","timestamp":"2026-09-15T09:02:22.817Z","sessionId":"99a4d3be-..."}
{"type":"queue-operation","operation":"dequeue","timestamp":"2026-09-15T09:02:22.819Z","sessionId":"99a4d3be-..."}
```
(2ms apart), suggesting the same queued operation got picked up/processed into a distinct session rather than being deduplicated or routed to the single active session.

Despite the identical origin, the three sessions have **zero overlapping message UUIDs** afterward — each proceeded completely independently, each producing several hundred unique messages (927 / 678 / 577 unique message UUIDs respectively) over the next ~1-4 hours, i.e. three full independent agentic runs of what should have been one conversation.

**Impact**

This tripled token/session-quota consumption for a single user request — in my case this one duplication alone accounted for roughly a third of my 5-hour rate limit window.

**Ask**

Could you look into the session dispatch/queue logic in the VS Code extension for a race condition where one submitted prompt (same `promptId`) gets delivered to multiple newly-created sessions instead of exactly one? Happy to share the raw JSONL lines (sanitized) if useful — I've kept the actual task content out of this report since it involved private repository work, but the structural/metadata evidence above (uuids, timestamps, queue trace) should be reproducible-looking without needing that content.

Begin with the VS Code extension's `claude-vscode` entrypoint and trace the session dispatch and queue-operation path using the supplied promptId and session trace evidence. Reproduce or instrument the enqueue/dequeue race, then add regression coverage showing that one promptId creates and reaches exactly one session without duplicate runs.

Written by the indexing model from the issue text.