[docs,hparams] docs: sync agent harness (.agents/.cursor/guidance) with implementation

[Jayce-Ping]

Summary

Resyncs the AI-agent harness with the current src/flow_factory/ implementation after plugin growth (now 9 trainers, 14 model adapters, 13 reward models). A full audit found ~40 discrepancies; this PR fixes registry drift, wrong config/API facts, and broken cross-references. Almost entirely docs/harness; one zero-risk type-hint correction in hparams/model_args.py.

Registry drift

• architecture.md / AGENTS.md: add diffusion-opd trainer, clap/imagebind/geneval rewards, Bagel/LTX2 models; fix RationalRewardsT2IRewardModel / RationalRewardsEditRewardModel class names.

Wrong facts that could mislead agents or break YAML

• guidance/algorithms.md: train.dynamics_type/train_steps/num_train_steps -> scheduler.dynamics_type/sde_steps/num_sde_steps (verified against examples/).
• guidance/workflow.md: real sample() (delegates to generate_samples) and compute_advantages() snippets (removed nonexistent compute_advantage_weighted_sum); fixed prompt_embedds typo.
• guidance/rewards.md: fixed trainer/grpo path, GenEval metadata-JSON convention, added audio param, added ocr/clap/imagebind.
• skills: model_path -> model_name_or_path; target_module_map -> config-driven + default_target_modules; data.dataset -> data.datasets; rewards single-dict -> list; removed evaluate() skeleton.
• constraints.md Optimize Groupwise Reward computation #11: evaluate() is concrete (not abstract).

Cross-reference rot

• constraints.md index extended to #28-29; philosophy.md #27 ref fixed and "FSDP" -> Accelerate (DDP/DeepSpeed ZeRO-1-2/FSDP).
• topics/samplers.md: corrected _resolve_sampler_type + AdvantageProcessor group_distributed path; parity_testing.md set_timesteps -> set_scheduler_timesteps; brittle line numbers -> symbol refs.
• .cursor/rules/skills-reusable-only.mdc: register skills in AGENTS.md; CLAUDE.md now imports @AGENTS.md so it can't drift to a stale subset.

Code (1 line of intent)

• hparams/model_args.py: model_type Literal now matches the 14 registry keys. Zero-risk — ArgABC is a plain dataclass with no Literal enforcement (examples already use sd3-5/bagel/ltx2_*).

Test plan

• Registry tables in docs re-grepped against trainers/, models/, rewards/ registries.
• model_args.py parses (AST) and has no linter errors.
• Re-grep confirms no stale tokens remain (compute_advantage_weighted_sum, trainer/grpo, prompt_embedds, bare train_steps, set_timesteps().
• Docs-only render check on GitHub.

Notes

• black --check/isort --check fail on src/flow_factory/hparams/model_args.py, but this is pre-existing (the whole file predates formatting); avoided a full-file reformat to keep this PR scoped. Can format separately if desired.
• Out of scope (code bug, not harness): my_reward_remote.py docstrings / reward_server/example_server.py reference flow_factory.rewards.remote instead of my_reward_remote.

Made with Cursor