Fix primary model detection and add advisory hook

[cri5Castro]

Outcome

Sol Advisor now verifies the active primary thread as exact gpt-5.6-sol / high before orchestration without sending valid root sessions through the subagent-only metadata path. The inspector reads the latest turn context, so an in-thread model change takes effect. Configured defaults alone do not count as active-session proof.

What changed

• Add backward-compatible --primary inspection for one exact root rollout while preserving auxiliary validation.
• Apply the primary evidence order: complete host metadata, local runtime inspection, then explicit same-thread confirmation after remediation.
• Require SELECTIVE ROUTE before local inspection and disclose when the gate relies on user attestation.
• Add the experimental $sol-advisor:configure-model-hook command and a model-only UserPromptSubmit advisory hook.
• Keep the hook non-blocking: it does not change models, prove reasoning effort, or authorize orchestration.
• Document TUI, CLI startup, and App Server model-selection paths.
• Bump the plugin release to 0.6.1 while preserving unrelated skill and manifest wording.

Verification

• sh plugins/sol-advisor/scripts/verify.sh
• Validate both Sol Advisor skills with quick_validate.py
• Validate the plugin manifest with validate_plugin.py
• git diff --check

The verifier covers latest-turn selection, exact mismatch reporting, malformed or ambiguous runtime metadata, unchanged auxiliary-role validation, hook allowlisting, unrelated prompts, and absence of reasoning-effort claims.

Trade-off

The bundled hook remains experimental and advisory because stable hook input exposes the active model but not reasoning effort. The primary runtime inspector remains the automatic authority for both fields, with same-thread user confirmation retained as the explicit fallback.