Phase 3: Flip shipped defaults to Anthropic + Haiku (all roles)

[neuromechanist]

Part of epic #155.

Goal: Flip HEDit's shipped defaults to native Anthropic with Haiku 4.5 for every role; retire the qwen/alibaba eval+vision defaults. OpenRouter/Ollama stay available as opt-in.

Scope

• DEFAULT_* model constants -> claude-haiku-4-5 for annotation, evaluation, assessment, feedback, keyword, and vision (src/cli/config.py:38-49; ModelsConfig :59-77). DEFAULT_PROVIDER -> anthropic.
• Native model-id namespace mapping (anthropic/claude-haiku-4.5 -> claude-haiku-4-5); per-provider handling of CACHEABLE_MODELS/OPENROUTER_MODELS.
• Reasoning-knob translation: OpenRouter reasoning.enabled=false -> native thinking (omit to disable; enable with a budget for annotation if we keep the Disable reasoning on non-annotation workflow LLMs (keyword, eval, feedback, assess) #150 behavior).
• Ignore the OpenRouter-only provider={"only":[...]} routing knob on the anthropic path.
• Per-provider credentials; migrate a flat openrouter_api_key in-place on load (config-format migration, not a runtime compat shim).
• CLI init/config flags (--provider, --base-url, --workspace-id); help text; .env.example + docs (Anthropic primary, OpenRouter/qwen as alternative).

Definition of Done

• Fresh install with only ANTHROPIC_API_KEY runs annotation + eval + vision end-to-end on Haiku, no OpenRouter/qwen involved.
• --provider openrouter and LLM_PROVIDER=ollama still work.
• Existing flat openrouter_api_key configs load without error (auto-migrated).
• Labeled breaking change; migration note in the PR and docs.

[neuromechanist]

Delivered by #162 (merged into develop 2026-08-19).

Shipped defaults are now claude-haiku-4-5 for annotation, evaluation, assessment,
feedback, keyword extraction, and vision; the qwen/alibaba eval and vision defaults are
gone, and Claude's native multimodality replaced the separate vision vendor. Legacy
OpenRouter-style ids (anthropic/claude-haiku-4.5) are accepted as aliases, and a flat
openrouter_api_key left in a credentials file is detected and reported instead of
failing obscurely.

Deviation from this phase's scope: OpenRouter and Ollama were removed rather than kept
as opt-in providers, so --provider flags do not exist. A modular
OpenAI/Anthropic-compatible endpoint is tracked in #163.

Reasoning-knob translation landed with a caveat now worth stating explicitly: on
Haiku 4.5 extended thinking is off unless budget_tokens is passed, so annotation
currently runs with no thinking on the default model, and disable_reasoning=True is a
no-op there. It maps to thinking: disabled on Sonnet 5, where thinking is on by
default. See #150 and #154.