Respect the project's model setting in prologue interpret

[woksin]

Fixed

• cratis prologue interpret no longer sends capture evidence to a machine-wide model when the project's cratis-prologue.json sets llm.enabled: false; --no-llm forces local, heuristics-only interpretation; and the command names the provider, destination and where its configuration came from before sending anything, rejecting an endpoint it cannot name (cratis prologue interpret uses the global language model when the local configuration disables it, and offers no way to turn the model off #241)

[woksin]

Reviewer notes (not part of the release note).

• Precedence: --no-llm overrides everything; an explicit local llm.enabled: false blocks the machine-wide fallback; an absent local setting keeps the fallback; an explicitly enabled local configuration resolves as before. The notice goes to stderr before the session starts, and the destination it names is the one the client uses (including ANTHROPIC_BASE_URL). An endpoint that is not an absolute URL is rejected before any client is created. The old LlmOptionsResolver.Resolve overload, which could not tell an absent setting from false, is removed (no production callers).
• Review: cross-provider deep review (Opus 5.5 + GPT-6 Astra) passed with no blocking findings; its three advisories (Anthropic base-URL disclosure, scheme-less endpoints, the ambiguous overload) are fixed in a770804.
• Local verification: restore, Release build with zero warnings, full dotnet test --configuration Release including integration, changed-file format and markdownlint — all exit 0 on 445658a; after the advisory fixes: Release build, CLI specs (3130 passed, 6 Windows-only skips), templating and conformance specs, format and markdownlint. Other-OS lanes run in CI only.