Fix lifecycle CI: replace defunct Anthropic model IDs and honor CODEFRAME_LIFECYCLE_MODEL

[Copilot]

The "Lifecycle (cli / haiku)" CI job was failing with anthropic.NotFoundError: model: claude-sonnet-4-20250514 — the early Claude 4.0 snapshot IDs are gone, and CODEFRAME_LIFECYCLE_MODEL=haiku set by the workflow was never actually wired into the agent's model selection.

Changes

Invalid model IDs (codeframe/adapters/llm/base.py, codeframe/core/adapters/streaming_chat.py)

• Replace claude-sonnet-4-20250514 → claude-sonnet-4-5 for planning/execution/correction/supervision defaults
• Replace claude-3-5-haiku-20241022 → claude-haiku-4-5 for generation default
• Update _DEFAULT_MODEL and pricing map in streaming_chat.py to match

CODEFRAME_LIFECYCLE_MODEL was silently ignored (tests/lifecycle/conftest.py)

• ModelSelector reads CODEFRAME_PLANNING_MODEL, CODEFRAME_EXECUTION_MODEL, etc. — not CODEFRAME_LIFECYCLE_MODEL. The cf subprocess inherited the alias (haiku/sonnet) but it had no effect, so the agent always fell through to the broken sonnet default.
• Added _lifecycle_model_env() that translates the alias to all five per-purpose env vars and merged them into every cf subprocess call:

_LIFECYCLE_MODEL_MAP = {
    "haiku": "claude-haiku-4-5",
    "sonnet": "claude-sonnet-4-5",
}

def _lifecycle_model_env() -> dict[str, str]:
    alias = os.getenv("CODEFRAME_LIFECYCLE_MODEL", "haiku")
    model = _LIFECYCLE_MODEL_MAP.get(alias, "claude-haiku-4-5")
    return {
        "CODEFRAME_PLANNING_MODEL": model,
        "CODEFRAME_EXECUTION_MODEL": model,
        "CODEFRAME_GENERATION_MODEL": model,
        "CODEFRAME_CORRECTION_MODEL": model,
        "CODEFRAME_SUPERVISION_MODEL": model,
    }