fix(embedding): bound vectorizer text input

[r266-tech]

Summary

Fixes #3010.

Collection vectorization currently forwards the full text field to the provider. If a file is larger than the embedding model accepts, the provider rejects the request and the file never gets a vector.

This PR applies the same configured input-bound idea at the VectorizerAdapter boundary:

• reads max_input_tokens from the vectorizer config when present
• bounds only provider-facing text before vectorize_document
• leaves the raw record data untouched, so stored fields/full-text indexing still keep the original content
• preserves image/video payloads
• treats missing/disabled/malformed limits as no limit, with warnings for malformed values
• fails closed before the provider call if the configured truncation helper is unavailable or raises

One tradeoff: dense and sparse provider generation share one text field in the current vectorizer API, so a configured cap bounds that shared provider input. The raw content field is still persisted unchanged.

Tests

• ruff check openviking/storage/vectordb/vectorize/vectorizer.py tests/unit/test_vectorize_file_strategy.py
• ruff format --check openviking/storage/vectordb/vectorize/vectorizer.py tests/unit/test_vectorize_file_strategy.py
• python3 -m py_compile openviking/storage/vectordb/vectorize/vectorizer.py tests/unit/test_vectorize_file_strategy.py
• direct adapter smoke check for configured cap behavior

I could not run PYTHONPATH=. pytest tests/unit/test_vectorize_file_strategy.py -q -o addopts='' in this unsynced local environment because collection imports require declared project dependencies such as json_repair that are not installed here.

Codex review: final pass approved after reworking config handling, fail-closed truncation paths, vectorize_one coverage, and media-field preservation.

[qin-ctx]

fixed by #3024