AgentSession.dispose() removes the host's event listener via _disconnectFromAgent() but does not abort any in-flight LLM HTTP request. Callers that dispose a session mid-stream (every switchSession / newSession / fork / clone via teardownCurrent) leave the previous LLM call running in the background:

• Orphaned TCP socket to the LLM provider — stays ESTABLISHED until the provider responds, often minutes for slow-stream APIs.
• Provider billing increments for the orphan call whose output is discarded.
• External observers see the process at 0% CPU with one ESTABLISHED HTTPS socket and no events flowing, indistinguishable from a wedged process.

Adds a synchronous this.agent.abort() at the top of dispose(), wrapped in try/catch so dispose remains exception-safe. AbortController.abort() is synchronous: the signal trips immediately even though the fetch rejection lands later. Keeps dispose() synchronous so existing callers don't need to be updated.

Discovered by PIRA when a session change against an already-loaded session triggered the orphan path, surfacing as 15-20 minute apparent hangs after any host-driven session change while a stream was in flight. PIRA shipped a host-side workaround (skip the redundant switch_session RPC when the bridge already has the session loaded) but the underlying dispose-without-abort behaviour bites any host that legitimately tears down a session mid-stream.

Recent in-the-wild repro: captured today (2026-05-26) — bridge sat wedged for 25 min after switch_session mid-turn. Trace tagged switch_session-teardown + subscriberless + llm-http-hang by our analyzer. Confirms the failure mode is legitimate.

Tests:

• agent-session-dispose-aborts.test.ts (4 cases): signal trips on dispose mid-stream, agent.abort called exactly once, exception-safe when agent.abort throws, idempotent on re-dispose.

Open question for review: dispose() only aborts the main agent. The session has several other AbortControllers (compaction, autoCompaction, branchSummary, retry, bash) that aren't tripped here. Recommend a follow-up PR (or expanding this one) to abort those too in dispose(), same orphan class, same fix pattern.