fix(daemon): log effective model capabilities with provenance

[Aaronontheweb]

Summary

The startup model-capability log reported the detector's raw output rather than the effective resolved capabilities. When an operator sets an InputModalities override on a provider that reports a context window but no modality metadata (e.g. vLLM), the detector returns a non-null partial result, so the log took the "auto-detected" branch and printed input=unknown — even though ModelCapabilityResolution had correctly applied the override. This made an applied override look ignored and sent at least one operator down a debugging rabbit hole.

Change

• Log the resolved (effective) capability values instead of the detector output, annotated with per-field provenance (configured / detected / default). Precedence mirrors ModelCapabilityResolution: configured override > detected > default, so an applied override can no longer be misreported.
• Add a one-line nudge — fired only when input modalities fall through to the default — pointing operators at Models:Main:InputModalities to enable vision on multimodal models behind metadata-less providers.

Example output now:

Resolved model capabilities for nvidia/Qwen3.6-27B-NVFP4: input=Text, Image (configured), output=Text (default), context_window=262144 (configured)

Related to the operator-experience side of #1127 (inline overrides on Models.Main were being wiped by doctor --fix; separately, when present they were being misreported by this log).

Test plan

• dotnet build src/Netclaw.Daemon clean (0 warnings)
• dotnet slopwatch analyze — 0 issues
• Add-FileHeaders.ps1 -Verify — all headers present
• ModelCapabilityResolution tests 6/6 pass (resolution precedence this log mirrors is already covered)

No log-string assertion test added: log output is brittle to assert on and low-value per the repo testing guidelines; the underlying precedence is covered by the resolver tests.

[Aaronontheweb]

LGTM