docs(config): state why Claude's rows have no long tiers, and stop implying a pending gap

[argszero]

Summary

The catalogue comment that explains why some models have no [[models.tiers]] said, in one sentence, that every unfilled row is a "the vendor has not published the number" case that is still pending ("未填的一律是「官网没给这个数」… 待补"). Two parts of that do not hold today.

Claude is not an unpublished number. Its pricing page has a Long context pricing section, and it states the rule rather than omitting it:

Claude 4.6 and later models (except Claude Haiku 5.5) and Claude Mythos Preview include the full 1M token context window at standard pricing. (A 900k-token request is billed at the same per-token rate as a 9k-token request.)

Both rows in this catalogue (claude-opus-5, claude-sonnet-5.5) are "4.6 and later" and neither is Haiku, so leaving them without tiers is the correct call — but the reason in the comment was "the pricing page has no long-context section" (「价格页无长档一段」), which is false. That matters because the note is the contract a future maintainer reads: as written it says "go find Claude's long-context numbers and fill them", inviting tiers the vendor never charges for.

The universal does not hold either. The unfilled families each have their own reason: the qwen3.7-flash tier table publishes only input/output columns, so adding tiers would drop its cache-hit price to zero (an undercharge); the OpenAI rows do have a short/long split but its boundary token count is not pinned down.

Changes

• config/config.example.toml — comment only (+9 / −3):
  • keeps the filled roster as it was (minimax-m3 >512k; gemini-3.1-pro / grok-4.6 / grok-4.7 >200k — re-checked against the file in this PR);
  • replaces "未填的一律是…待补" with one reason per family, so no row is implicitly promised as a pending gap;
  • records the rule that applies if Claude Haiku 5.5 is ever added: it is priced by prompt length with the boundary at > 100,000 tokens (counting cache reads and writes), i.e. from_tokens = 100001.

No behaviour, data or configuration-value change: the file parses to the same [[models]] set, and the affected lines are comments.

Tests

• cargo test — 579 passed / 0 failed (same count as before: comment-only)
• cargo fmt --check — clean
• No new tests: the change is prose; the numbers it describes are owned by the vendor page, not by this repo

Checklist

• Branch name follows the convention (docs/)
• Commit message follows Conventional Commits
• Single responsibility, minimal change

Source

The finding was measured on 2026-10-18 while auditing whether the reasons this catalogue gives for its own empty tiers still hold. The vendor rule above was re-fetched verbatim from https://platform.claude.com/llms-full.txt (which mirrors docs/en/about-claude/pricing) for this PR.