- Add plugin schema types and runtime discovery for meta/filter plugins
- Rewrite --generate-config to use schema system instead of hardcoded types
- Add Settings::validate_config() for startup validation
- Cache tokenizer instances via static Lazy to avoid repeated BPE loading
- Add split_by_token_iter() and count_bounded() to Tokenizer
- Fix double-counting bug in TokensMetaPlugin when buffer < max_buffer_size
- Eliminate unnecessary allocations in token count methods
- Refactor token filters: remove Option<Tokenizer>, use iterator API
- Fix TailTokensFilter correctness: unbounded buffer instead of ring buffer
- Add encoding option to all token filters
- Add description() to MetaPlugin and FilterPlugin traits
- Fix unused_mut warning in compression engine (feature-gated code)
Co-Authored-By: code-review-bot <noreply@anthropic.com>
Add tiktoken-based token counting via new 'tokens' feature flag.
New components:
- Shared tokenizer module wrapping tiktoken CoreBPE (cl100k_base, o200k_base)
- TokensMetaPlugin: streaming token counter, tokenizes each chunk independently
- head_tokens(N): stream first N tokens, split at exact boundary when mid-chunk
- skip_tokens(N): skip first N tokens, stream the rest
- tail_tokens(N): bounded ring buffer (~16KB), outputs last N tokens at finalize
All filters are fully streaming — no full-stream buffering.
Meta plugin accuracy: exact for normal text, ±1-2 tokens if long whitespace
sequence spans a chunk boundary.
Also: add 'client' and 'tokens' to default features, add curl to Dockerfile builder stage.
