fix(custom-profile): omit reasoning_effort="none" for remote endpoints

[ajzrva-sys]

Fixes #65233
Related: #59660

Problem

CustomProfile.build_api_kwargs_extras() unconditionally sent reasoning_effort="none" + think=False when reasoning was disabled. Remote OpenAI-compatible APIs (ofox, Volcengine ARK, Doubao, etc.) reject reasoning_effort="none" as invalid — HTTP 400.

Fix

Only emit both fields for local endpoints. Remote endpoints get neither — server default applies.

Local endpoint detection (proper, not substring matching)

1. urlparse for hostname extraction — avoids false positives like https://localhost.example.com being treated as local
2. Exact hostname matches: localhost, ip6-localhost, ip6-loopback, 0.0.0.0, ::, ::1
3. IPv6 loopback support: [::1], [::] recognized
4. RFC1918 private/LAN ranges: 192.168.x.x, 10.x.x.x, 172.16-31.x.x via ipaddress module
5. Link-local addresses: 169.254.x.x
6. mDNS .local domains: Bonjour/Avahi (Ollama on other Macs, AirDrop)
7. Fallback: no base_url → default to local (backward compat with legacy tests that don’t pass base_url)

Wire behavior

Scenario	Fields emitted
disabled + local	reasoning_effort="none", think=False
disabled + remote	nothing
effort="none" + local	reasoning_effort="none", think=False
effort="none" + remote	nothing
effort=low/medium/high/max/xhigh	reasoning_effort=<effort> (passed through)
enabled, no effort	nothing

Testing

• All 13 existing tests pass
• 4 new tests added: remote-local split for both enabled=False and effort="none" paths
• Valid effort levels (low/medium/high/max/xhigh) still pass through verbatim

[teknium1]

Thanks for isolating the current custom-profile wire regression; current main does unconditionally emit both disable fields at plugins/model-providers/custom/__init__.py:57-65.

Problems

• plugins/model-providers/custom/__init__.py:126 uses endpoint locality as the capability test. The custom provider is documented to target any OpenAI-compatible endpoint (website/docs/integrations/providers.md:570), so locality does not establish whether an endpoint accepts reasoning_effort="none" or think. It also changes an explicit disabled setting into server-default behavior for every remote endpoint (PR lines 106-108), matching the concern raised in #65233's discussion.
• The new remote assertions invoke the profile directly (tests/plugins/model_providers/test_custom_profile.py:73-78). The production wire path is assembled in agent/transports/chat_completions.py:582-611; add a transport-level regression test for the final kwargs.

Suggested changes

• Model disabling support as an explicit custom-endpoint capability/profile decision rather than a hostname classification, retaining the Ollama-specific path where it is known to be required.
• Cover both supported-disable and unsupported-disable endpoint cases through transport.build_kwargs.

Automated hermes-sweeper review.

[teknium1]

This branch infers API parameter support from network locality, but provider=custom is documented for arbitrary OpenAI-compatible endpoints (website/docs/integrations/providers.md:570). A local proxy can reject these fields and a remote endpoint can support them; please select this behavior through an explicit endpoint/provider capability instead, so enabled: false is not silently converted to an endpoint default based only on its hostname.