Follow-ups from the 2026-11 token-count update #19

This issue tracks follow-ups from the token-count update (PRs #10, #14, #15, #16 and their successors). I add an item when the code it concerns is on main. Each item names the PR that fixes it or the trigger that starts it.

From #10 (Claude encodings, merged)

• Map Claude models to claude_v3, claude_v4_7, claude_v4_8 and claude_v5 in models.json, with per-model frames and tool overhead. The Claude models PR (branch feat/claude-models) does this.
• Apply the claude_v5 trailing-whitespace rule only at the end of a user message and the system prompt in count(). The encoding trims every text it counts. The Claude models PR does this.
• A final single space after U+302A to U+302F counts 1 too many on claude_v3, claude_v4_7 and claude_v4_8. Unicode 16 letters U+A7CC, U+A7CD and U+A7DA count about 1 off. Both gaps come from ctok's reconstructed vocabulary and are documented in the README. Trigger: a ctok upstream fix, then port it.

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high