Sanitize embedding text and surface request errors

[jordanpartridge]

Problem

OllamaEmbeddings/OpenAiEmbeddings swallowed every request failure (catch (RequestException) → return []), so consumers only ever saw a vague "failed to generate embedding" with no cause. The actual failure being masked: bge-large has a 512-token BERT context, and unicode-dense content (emoji, box-drawing trees, smart punctuation) byte-fallbacks into so many tokens that Ollama's truncation miscounts and rejects the request with HTTP 400 the input length exceeds the context length — while plain prose of 18KB embeds fine. Those characters carry no retrieval signal anyway.

Changes

• EmbeddingTextSanitizer — normalizes text before vectorization: strips emoji/pictographs/dingbats/box-drawing/arrow blocks, deletes zero-width chars and variation selectors, maps smart punctuation to ASCII, collapses leftover spacing. Lossy by design — only the embed request sees the sanitized text, stored content is untouched.
• OllamaEmbeddings — sanitizes by default (sanitize: true constructor flag to opt out).
• OpenAiEmbeddings — same flag, default off (8k context, robust tokenizer).
• EmbeddingRequestException — replaces the silent empty-array return on request failure; carries model, HTTP status, and the provider's error message (OpenAI nested error.message, Ollama flat error, raw body fallback for non-JSON responses).

Verification

• Full suite passes at 100% coverage; pint + rector clean.
• End-to-end against production Ollama (odin) with a real document that previously failed:
  • default: returns a 1024-dim vector
  • sanitize: false: throws Embedding request failed for model [bge-large] (HTTP 400): the input length exceeds the context length

Breaking change

embedBatch()/embed() now throw on request failure instead of returning empty arrays. Callers that branched on [] === $vector (e.g. conduit-ui/knowledge's QdrantService) will get the descriptive exception directly — which is the point — but should be updated when bumping.

[lexi-chief-of-staff]

Ship it.

• Breaking change is intentional and documented; callers that previously tested [] === $vector will now surface real errors.
• Sanitizer applied only to the embed path, stored content untouched.
• OpenAI keeps sanitize: false default (8k+ context), Ollama defaults true (512-token BERT).
• Mock tests cover the new exception path, non-JSON fallback, and sanitization toggle.

[lexi-chief-of-staff]

Ship it.

• Breaking change on failure path is explicit and documented; callers now get model + status + provider message instead of swallowing into [].
• Sanitizer default-true only on Ollama (correct default given 512-token constraint); OpenAI stays opt-in.
• Regression test in EmbeddingsTest asserts the new exception path and the 400 context-length message.
• Dependency swap (nunomaduro/pao → laravel/pao, plus transitive upgrades) is isolated to composer.* and passes pint/rector.
• No style or layout issues on the new sanitizer or exception classes.