feat: optionally enforce playbook token budget

[Frankie-Xu]

Addresses the playbook-bloat concern in #26 with an opt-in budget guardrail. It does not implement DELETE/MERGE/UPDATE, replace the prompt loader in #38, change PR #43, or claim that a budget alone fixes validation oscillation.

This adds enforce_playbook_token_budget (default false) to ACE, ACEBatch and the three task entry points. When enabled, the existing count_tokens / cl100k_base counter measures the resulting playbook only. The curation boundary rejects an over-budget candidate before publication, returns the original playbook unchanged, does not consume next_global_id, and records structured diagnostics: status, base/candidate/budget token counts and reason. Bullet content is never silently truncated. Empty operations and exact-boundary candidates are covered.

Ordinary Curator curation attaches the decision to call_info; the ACEBatch reducer uses the same pure helper before assigning its playbook and ID and exposes the decision in per-sample tracking. The default path retains existing tuple/config behavior.

Tests cover exact-boundary acceptance, over-budget rejection, empty operations, default compatibility and opt-in config extraction, Curator structured rejection, reducer output/state preservation, and no ID consumption. Validation: 7 unittest tests passed, compileall and diff-check passed, and the same suite passed from a clean git archive. No LLM, API key, network request, benchmark, or upstream main push was used.