fix(reasoning): normalize custom-provider model ids for fallback detection

[Carry00]

Thinking Path

• Hermes WebUI should expose the reasoning-effort selector for thinking-capable models even when they are served through named custom:* providers.
• The current fallback path already handles slash-prefixed ids and a small set of hand-written bare-name prefixes.
• That approach fixed deepseek-v4-flash, but it still depends on exact separator shape and misses equivalent ids such as deepseek.v3.2 or deepseek_v4_flash.
• Custom API aggregators frequently rewrite model ids with dots, underscores, or extra vendor namespaces, so format-specific prefix checks are inherently brittle.
• This PR makes the fallback heuristic normalize non-slash model ids before matching them against reasoning-capable model families.
• The result is a more durable fix for custom-provider model naming drift, while also narrowing a known false-positive path in the old substring-based keyword fallback.

What Changed

• Added _reasoning_name_candidates() in api/config.py to generate normalized candidate names for heuristic reasoning-capability checks.
• Added _candidate_supports_reasoning() to match reasoning-capable model families by normalized tokens/family semantics instead of raw string shape.
• Replaced the previous custom-provider fallback block in _heuristic_reasoning_efforts() with candidate-based family matching.
• Kept the existing resolution order unchanged:
  • provider-specific authoritative paths first
  • models.dev metadata next
  • heuristic fallback last
• Added regression coverage for separator and namespace variants:
  • deepseek.v3.2
  • deepseek_v3_2
  • vendor.deepseek.v3.2
  • deepseek.v4-flash
  • deepseek_v4_flash
• Added negative coverage to prevent vendor-prefix keyword false positives:
  • thinkinghub.llama-3.1-70b
  • reasoninghub.llama-3.1-70b
• Updated CHANGELOG.md with the user-visible behavior change.

Why It Matters

• Named custom providers do not consistently preserve model-id separator style.
• The old heuristic assumed a few exact shapes (vendor/model, vendor.model, selected bare prefixes), so equivalent DeepSeek ids could silently hide the reasoning-effort selector.
• This PR improves resilience to naming variation without broadening the authoritative paths or changing built-in provider behavior.
• It also fixes an adjacent correctness issue where substring-based keyword matching could falsely enable reasoning for unrelated vendor prefixes like thinkinghub.*.

Verification

• Runtime verification via direct resolve_model_reasoning_efforts() calls:
  • deepseek.v3.2 -> returns full supported efforts
  • deepseek_v3_2 -> returns full supported efforts
  • vendor.deepseek.v3.2 -> returns full supported efforts
  • thinkinghub.llama-3.1-70b -> returns []
  • reasoninghub.llama-3.1-70b -> returns []
• python3 -m py_compile api/config.py
• VS Code diagnostics checked for:
  • api/config.py
  • tests/test_custom_provider_bare_model_reasoning.py
  • CHANGELOG.md

Risks / Follow-ups

• This remains a heuristic fallback for cases where no authoritative metadata exists, so future provider-specific naming schemes may still need family-rule expansion.
• A stronger long-term direction would be allowing explicit capability overrides for named custom providers/models, but that is intentionally out of scope here.
• Automated pytest execution could not be completed in this environment because pytest is not installed (python3 -m pytest -> No module named pytest).

Model Used

• GPT-5.4 via Trae coding agent
• Tooling used: repository search/read, patch editing, local Python runtime checks, diagnostics

[nesquena-hermes]

Shipped in v0.51.198 (stage-batch10, via #3350). Thanks @Carry00 — the token-aware normalization (vendor-namespace stripping + separator folding) cleanly fixes the reasoning-effort detection for custom:* providers. Opus + Codex both confirmed it only affects the custom-provider fallback heuristic path and doesn't regress existing slash/@-qualified ids. 🎉