feat: add per-phase Claude model configuration (claude_model + review_model)

[dwilberger]

Summary

• Add claude_model and review_model config parameters to control which Claude model is used for task execution vs review phases
• Add --claude-model and --review-model CLI flags with standard precedence (CLI > local config > global config > default)
• Create separate review executor when review_model differs from claude_model, used by all review phases (first review, fix loops, codex eval, finalize)
• Empty values preserve current behavior (no --model flag passed to Claude CLI)

Motivation

Review phases run 5-7 parallel Claude agents that can iterate multiple times. Using the same model (e.g., Opus) for both tasks and reviews is unnecessarily expensive. This change allows using a capable model for tasks while using a faster/cheaper model for reviews.

Usage

# in ~/.config/ralphex/config
claude_model = opus
review_model = sonnet

# or via CLI flags
ralphex --claude-model opus --review-model sonnet docs/plans/sprint-1.md

Test plan

• Config parsing tests: claude_model and review_model parsed from INI
• Merge logic tests: local overrides global, empty preserves existing
• Executor test: --model flag injected when Model is set, absent when empty
• Full build: go build ./... passes
• Full test suite: all pre-existing tests pass (failures are pre-existing Docker/permission issues)
• No secrets or env values in diff

Closes #271

[umputun]

nice feature, thx. couple things to address:

1. naming - claude_model is misleading, it only applies to task/planning phases, not to claude globally. task_model would be more accurate and symmetric with review_model. same for the CLI flag - --task-model / --review-model reads better than --claude-model / --review-model
2. config comment typo - pkg/config/defaults/config:33 says # claude_model = but should be # review_model =
3. custom wrappers - --model flag is silently ignored by custom providers (claude_command config). not a bug, but worth documenting that model config only applies with the default Claude CLI
4. missing processor test - the ReviewClaude executor separation logic in runner.go has no test coverage. existing tests only exercise the fallback path (ReviewClaude: nil). a test with distinct task/review mock executors verifying the review phase calls the right one would close this gap
5. no docs update - README, CLAUDE.md, and llms.txt document all config options but don't mention the new fields or CLI flags

the approach itself is clean - separate executor only when models differ, correct config merge chain, backward compatible. no scope creep, tests pass, linter clean.