v0.20.3

Fixed

Two observer crash modes and one upstream-pin bump — all aimed at the supervised observer surviving real-world macOS conditions.

• asyncio.get_event_loop() → get_running_loop() in the observer coroutine (memory/observer.py:212). On Python 3.14 the deprecated call has been observed raising NameError: name 'asyncio' is not defined from inside the coroutine launched by asyncio.run(self._run_async()), even though asyncio is imported at module scope and the same name resolves fine in run() a few lines up. get_running_loop() is the canonical API inside a running coroutine, dodges the deprecation, and stops the supervisor from chewing through max_restarts: 10 on every restart.
• fastembed cache pinned to ~/.cache/neo/fastembed/ (memory/store.py). The default cache lives under $TMPDIR/fastembed_cache/, which on macOS is /var/folders/<...>/T/ — periodically swept by the OS. After a sweep the in-cache manifest still points at a (now-deleted) model.onnx and TextEmbedding(...) raises ONNXRuntimeError ... NO_SUCHFILE, leaving FactStore silently running without embeddings until a human noticed. The new path survives reboots and tmp sweeps; on NO_SUCHFILE / missing-model.onnx the loader nukes the stale snapshot dir and retries once, so any future eviction self-heals.

Changed

• car-runtime pin bumped to >=0.18.0 for two upstream supervisor-reliability fixes that directly hit neo memory observer: spawn_supervision is now teardown-first with kill_on_drop, so a start issued during Backoff can no longer leave an orphaned child holding the agent's port; and restart backoff is exponential (base → ×2, capped at 60s) instead of flat, so a crash loop can no longer burn through max_restarts in under a minute. last_exit_code is also now cleared on successful (re)start, so neo memory observer status won't show a stale 1 next to a healthy running agent. No neo code changes required — the pin captures the upstream wins for new installs.