[nicoloboschi]

Fixes #1897

Problem

Model/connection initialization had no wall-clock cap. If embeddings, the cross-encoder, or LLM verification blocked — e.g. an offline HuggingFace download or an unreachable provider — asyncio.gather in MemoryEngine.initialize() (memory_engine.py:2313) never returned. The daemon got stuck in a third state: not started, not errored, just hung. From the caller's side the memory operation hangs forever with no way to recover.

The lazy reranker path CrossEncoderReranker.ensure_initialized() (reranking.py:156) had the same gap, surfacing the hang on the first recall instead of at startup.

Fix — fail fast, no silent fallback

• New static config HINDSIGHT_API_MODEL_INIT_TIMEOUT (default 300s — generous enough for legitimate first-time model downloads on slow connections).
• Wrap the startup gather and the lazy reranker init with asyncio.wait_for. On timeout, raise a RuntimeError that names the likely cause and points at the env var.
• No graceful degradation / RRF fallback — consistent with the prior fix(api-slim): gracefully handle model init failures #1879 discussion that silent degradation is the wrong approach.

Deliberately kept minimal-scope: did not add the diagnostic _embeddings_initialized / _reranker_initialized state fields the issue floated — the reranker already tracks _initialized and they aren't needed for the fix.

Note: env var follows the project's HINDSIGHT_API_* convention (the issue suggested HINDSIGHT_MODEL_INIT_TIMEOUT).

Tests

tests/test_model_init_timeout.py:

• timeout fires on a hung init → RuntimeError
• fast init completes and marks the reranker initialized
• default config value is 300s

Lint, format, ty type-check, and existing reranker-timeout tests all pass.