fix(embeddings): truncate oversized litellm-sdk inputs before embedding (#2501)

[nicoloboschi]

Fixes #2501 (partial — the embedding-overflow half).

Problem

A pinned mental model in delta refresh mode accumulates content that grows monotonically across refresh cycles. Once it exceeds the configured embedding model's fixed input-token limit (e.g. Bedrock Titan V2's hard 8,192-token cap), refresh_mental_model fails permanently — every retry hits the same ContextWindowExceededError because the stored content never shrinks. The only recovery today is manually calling POST .../mental-models/{id}/clear.

The embedding call chain (LiteLLMSDKEmbeddings.encode() → litellm.embedding()) had zero length check or truncation.

Fix

Add an opt-in HINDSIGHT_API_EMBEDDINGS_LITELLM_SDK_MAX_INPUT_TOKENS cap. When set, encode() truncates each input to that many cl100k_base tokens before calling litellm.embedding() — mirroring the existing reranker HINDSIGHT_API_RERANKER_LITELLM_MAX_TOKENS_PER_DOC / _truncate_to_tokens pattern the issue references.

• Truncation emits a log.warning naming the model and the largest original token count, so it's never silent.
• Off by default → no behavior change or data loss for large-context embedding models. Titan users set it to the model's real limit (e.g. 8192) with a little headroom (tiktoken only approximates the provider's tokenizer).
• Server-level static config (embeddings aren't per-bank overridable).

Scope

This addresses the embedding call overflow (gap #2 in the issue) — the direct cause of the permanent failure — per the "just truncate the vector indexing for now" steer. The deeper root cause (gap #1: unbounded delta-document growth in apply_operations()) is not addressed here and is worth a follow-up if we want to cap persisted document size.

Tests

tests/test_litellm_sdk_embeddings.py:

• oversized input is truncated to the cap and a warning is logged
• warning does not fire when inputs already fit
• text passes through verbatim when the cap is unset (default)

All 29 tests pass; ruff + ty + lint.sh clean.