Add agentic training data synthesizer

[tushuhei]

This PR introduces a three-stage agentic synthesis architecture to dynamically generate, align, and polish candidate sentences to supplement the training dataset.
This pipeline is highly experimental and only focused on Japanese phrase-based segmentation for now.

Pipeline Workflow

1. Intent Understanding Agent (extract_intent_target & parse_direct_input): Parses direct CLI patterns (--input="いよいよ/はじまる") or GitHub bug report descriptions (--issue=468) to identify targeted segments (expected_split) and confirm reproduction against live production models (ja.json).
2. Example Generator Agent (generate_oversample_candidates): Synthesizes varied natural language corpus sentences strictly enclosing the targeted break intervals.
3. Boundary Alignment & Linguistic Expert Review:
   • align_to_base_parser_splits: Retains existing parser segmentations across surrounding outside context to preserve baseline unigram training integrity while strictly changing only the target interval.
   • prune_linguistic_anomalies: Audits syntax fluency and polishes improper boundary fusions around independent vocabulary words without splitting across enclosing conversational quotation brackets (「」).

Usage

# Direct targeted string segmentation
uv run scripts/synthesize_samples.py --input="いよいよ/はじまる" --lang ja --outfile staging_ja.txt

# Automated bug parameters derived straight from a public issue report
uv run scripts/synthesize_samples.py --issue=468 --lang ja --outfile staging_468.txt

[kojiishi]

LGTM