Summary

Finishes the long-started transition from a raw byte file-size gate to a token-based limit for AI-read docs — using a no-auth, offline tokenizer (tiktoken) instead of atc (which needs ANTHROPIC_API_KEY). Both gates now coexist with mutually-exclusive coverage: token budgets for Markdown docs, byte limits for everything else.

Changes

• _token-limits.yml (new, reusable): counts tokens with public/offline tiktoken (pip install tiktoken, no secret), per-repo config in .token-limits.yaml; sparse-checkouts the shared counter (same pattern as the watchdog).
• scripts/check-token-limits.py (new): fnmatch budgets (most-restrictive match wins), exclude globs, skips non-token-gated + binary files; exit 1 on violation. A file is token-gated iff it matches a limits pattern.
• _file-size.yml: when a .token-limits.yaml is present, drops .md from its scan so Markdown docs are governed only by the token gate — no file is double-gated. Repos without .token-limits.yaml are unaffected.
• _ci-gate.yml: new token_limits input + Token Limits job wired into the Merge Gate needs/allowed-skips.

Config shape

# .token-limits.yaml
defaults: { max_tokens: 2000 }
exclude: ['TERRAFORM.md']      # machine-generated reference: neither gate
limits:
  AGENTS.md: 2000
  '*README.md': 1500
  '*.md': 2000                  # catch-all so all .md are budgeted

Test plan

• actionlint clean on all three workflows (shellcheck included — no SC2053/SC2254 suppressions; the partition uses an extension drop, not dynamic globs).
• Counter verified locally with real tiktoken (o200k_base): correct per-pattern budgets, exclude honored, non-.md skipped, exit 1 on over-budget.
• No secrets required anywhere.

First consumer: dryvist/tofu-unifi (follow-up PR) flips file_size-only → file_size + token_limits.