Add dry-run subcommand for token-usage measurement

[jehna]

Summary

• New humanify dry-run <file> [--context-size N] [--tokenizer cl100k_base|o200k_base] subcommand. Walks every identifier the real pipeline would walk, builds the exact (system, user, schema) triple the LLM renamer would send, tokenises it via tiktoken-rs, and reports calls plus per-call and aggregate token totals. No LLM, no API key.
• Extracts build_rename_prompt() out of LlmRenamer so the dry-run path tokenises the exact bytes the real path would have sent — no chance of drift between the two.
• Used to produce the numbers in v3 uses ~3× more tokens than v2 for the same input #749 (token-usage comparison vs. v2).

Example

$ humanify dry-run fixtures/example.min.js
humanify dry-run
tokenizer       cl100k_base
context_size    500
calls           6
system tokens   240
user tokens     822
schema tokens   234
total tokens    1296
avg total/call  216
avg user/call   137

Test plan

• cargo test --release (186 passing, no regressions)
• Manual run on fixtures/example.min.js
• Manual run on a synthetic shadowing fixture (numbers match v3 uses ~3× more tokens than v2 for the same input #749)
• Reviewer: try humanify dry-run on a real bundle and sanity-check the totals

---

Generated by Claude Code