[No.35] feat(doctor): add config doctor and dry-run

[AutumnCs]

Task issue: #203

What

Add a CI-friendly Config Doctor and dry-run entrypoint for Task 35:

relax-doctor [doctor options] -- <training arguments>

The command statically resolves training configuration and reports:

• merged and normalized configuration with conservative field provenance
• runtime role and placement-group topology
• GPU requirements and parallel-rank derivation
• sanitized expected launch command
• structured diagnostics with stable rule IDs and targeted fixes
• resolved, partial, or unresolved resolution status

It does not start Ray, Ray Serve, SGLang, CUDA/distributed workers, or a training loop.

Why

Relates to #203 and #86.

Many argument, path, dependency, resource, backend, and algorithm conflicts can be identified before GPU allocation. Doctor, Registry, Controller, and managed-teacher helpers consume the same side-effect-free Launch Plan primitives so the preview follows runtime role and GPU semantics instead of maintaining a Doctor-only topology.

How

• Reuse the real Relax, Megatron, and SGLang parser stack for deterministic static normalization and backend validation.
• Preserve the original parser error on fallback; partial output is diagnostic-only and omits authoritative topology, GPU, and rank plans.
• Reject unknown long/short options, malformed resources, booleans used as integers, zero-GPU model roles, and GPU-assigned CPU-only roles.
• Derive GRPO/PPO/SFT, debug, Hybrid, fully-async, true-on-policy, GenRM, and managed-teacher plans through shared pure functions.
• Parse a documented safe Bash subset without executing scripts; unresolved source files, required variables, command substitution, control flow, malformed syntax, or missing training commands fail closed.
• Support structured config, generalized data paths, text/JSON output, --output, --strict-warnings, and exit codes 0/1/2.
• Recursively redact key/token/password/secret/credential values from argv, config, runtime env, commands, parser errors, rule failures, text, and JSON.
• Provide an extensible rule registry, more than 15 error classes, a fixture library, and Chinese/English documentation.

Testing

Environment:

• Linux, Python 3.12.13
• rebased onto main@ce650113e01b2c0b1dabbfe3f35cdffd7f3be67d
• CPU-only static validation; no GPU worker was allocated or started

export PYTHONPATH=/path/to/Relax:/path/to/Megatron-LM

python -m pytest -q tests/doctor tests/core/test_launch_plan.py \
  tests/utils/test_arguments_opd_teacher_colocate.py
# 114 passed

python -m pytest -q tests --tb=short
# 1011 passed, 14 skipped

pre-commit run --all-files
# all hooks passed

npx vitepress build docs
# build complete

python -m compileall -q relax
git diff --check
# passed

Warnings are existing third-party deprecation/capability warnings; no failed test was skipped. GitHub CI should be evaluated again after the rebased commits are pushed.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation update
• Refactoring to share static launch planning
• Performance improvement
• CI/CD or build change

Risk & Rollback

• --script intentionally supports only a safe static subset and fails closed outside that subset.
• Missing parser dependencies produce a non-success partial report, never a false final configuration.
• Runtime-facing changes are limited to shared static planning/validation helpers and retain existing compatibility interfaces.
• Reverting this PR removes the entrypoint and shared planning refactor.

Checklist

• Diff is limited to Task 35 implementation, tests, fixtures, and documentation
• Full pre-commit passes
• Related and full repository tests pass
• VitePress documentation build passes
• No datasets, checkpoints, private machine paths, or real credentials are included
• GitHub CI passes on the latest pushed commits
• Maintainer review comments are addressed