[P2.27] Route remaining direct provider constructions through the shared llm_resolution chain

[frankbria]

Severity: medium | Category: correctness | Follow-up to: #768 (PR #860)

Problem

#768 introduced codeframe/core/llm_resolution.py (flag → CODEFRAME_LLM_PROVIDER → .codeframe/config.yaml → anthropic) and routed the CLI pre-flight, runtime.execute_agent, prd_v2, and discovery_v2 through it. Four paths still construct providers directly and ignore the chain:

• codeframe/core/conductor.py:125 — SupervisorResolver.llm uses bare get_provider() (anthropic). Concrete failure: cf work batch run --strategy auto --llm-provider ollama passes pre-flight, but blocker auto-resolution downstream still demands Anthropic.
• codeframe/core/dependency_analyzer.py:229 — AnthropicProvider(api_key=...) hardcoded
• codeframe/core/prd_discovery.py:250 — AnthropicProvider(api_key=key) hardcoded
• codeframe/core/adapters/streaming_chat.py:189 — AnthropicProvider(api_key=key) hardcoded

Acceptance criteria

Each site resolves its provider via resolve_llm_settings/create_provider (or explicitly documents why it is Anthropic-only), with a regression test per site mirroring tests/ui/test_discovery_generate_tasks.py.

Dependencies

None (builds on #768, merged)

---

_{Spun off from the #768 internal review (suggestion tier). Atomic: one developer, one session.}