Why:
- Indexing against Ollama WordPiece-based embedding models can exceed model context limits because chunk sizing used only cl100k_base token counts.
- Ollama error payloads using "input length exceeds the context length" were not consistently classified as token-limit failures.

What:
- Added tokenizer strategies (tiktoken, character, simple) and strategy-based default tokenizer factory.
- Added ACI_TOKENIZER config/env support and wired tokenizer selection into service initialization so chunking and summary generation share the chosen tokenizer.
- Expanded token-limit error detection to include context-length style 400 responses.
- Updated unit/property tests for tokenizer strategy selection, offline-safe tokenizer tests, config strategy generation, and token-limit message coverage.

Test:
- uv run pytest tests/unit/test_tokenizer.py tests/property/test_embedding_client_properties.py tests/property/test_config_properties.py -q (pass)
- uv run ruff check src tests (pass)
- uv run pytest tests/ -v --tb=short -q --durations=10 --maxfail=1 (fails in this environment due proxy blocking tiktoken encoding download)
- uv run mypy src --ignore-missing-imports --no-error-summary (existing repo-wide mypy errors)