fix(reasoning): add GLM and Claude family-before-version names to heuristic family list

[Carry00]

What

PR #3327 added dot-separator normalization so custom-provider model IDs like deepseek.v3.2 correctly surface the reasoning-effort selector. However, two model families were left out of _candidate_supports_reasoning and remain broken after that fix.

GLM (ZAI / Zhipu AI) — zai.glm-5, zai.glm-4.7, glm-5, etc.
All modern GLM generations (4.5+) have reasoning: true in models.dev, but glm was never added to the heuristic family list. The dot-stripping from #3327 correctly surfaces glm-5 as a candidate — it just has nowhere to match.

Claude family-before-version naming — claude-opus-4-8, claude-haiku-4-5, anthropic.claude-opus-4-8
The existing check is startswith("claude-3", "claude-4"), which matches the old claude-3-5-sonnet-... format. The newer naming convention puts the family before the version (claude-opus-4-8), so it starts with claude-opus, not claude-4, and silently misses.

Changes

• api/config.py — two additions to _candidate_supports_reasoning:
  • glm- prefix for the GLM family
  • claude- + {3, 4} token intersection for the new Claude naming convention
• tests/test_custom_provider_bare_model_reasoning.py — 11 new parametrized cases (28 total, all pass)
• CHANGELOG.md — user-visible entry

Verification

28 passed in 3.74s

Models now correctly detected (were all [] before):

Model ID	Before	After
zai.glm-5	[]	full efforts
zai.glm-4.7	[]	full efforts
anthropic.claude-opus-4-8	[]	full efforts
anthropic.claude-haiku-4-5	[]	full efforts
deepseek.v3.2	✓ (from #3327)	✓

Non-reasoning models unchanged (meta.llama-3.1-70b, thinkinghub.* → []).

---

Note on the underlying design

This is a targeted fix but also a band-aid. The heuristic family list (_candidate_supports_reasoning) will need a new entry every time a provider uses an unfamiliar naming convention — GLM and Claude are the two that surfaced this week, but there will be more.

A more durable solution would be to allow per-model capability overrides in the custom provider config, so users (or aggregator presets) can declare reasoning: true without requiring a code change. Happy to prototype that if maintainers think it's worth pursuing — just want to flag the root cause before closing it with another patch.

[nesquena-hermes]

Closing as superseded by #3379, shipped in v0.51.211 (Release GE). Thanks @Carry00 — your additions (GLM models + Claude family-before-version naming like claude-opus-4-8/claude-haiku-4-5) are both fully covered by #3379's broader family-based rewrite of _candidate_supports_reasoning. I verified your exact target cases (zai.glm-4.7, glm-5, anthropic.claude-opus-4-8, claude-haiku-4-5) all resolve correctly with the merged change. Appreciate the fix — it pointed at the same gap.