feat(sd3): dump-verified boundary input dtypes + fp32 OCR recipe

[Rockdu]

Stacked on #91 (feat/fsdp-precision-control) — only the last 2 commits are new; will rebase onto main once #91 merges.

What

• configs/sd3.py: SD3 opts into the boundary input dtype policy — {"latents": "default", "cond": None, "timestep": "fp32"} (base default is full passthrough after [Do Not Merge] refactor(fsdp)(4/5): per-parameter weight-precision control for FSDP2 #91).
• scripts/run-diffusion-grpo-sd3-ocr-sglang.sh: --precision-default-dtype fp16 → fp32.
• arguments.py + sglang_diffusion_engine.py: forward --sglang-text-encoder-precisions to the rollout engine (same explicit-forward pattern as dit_precision); the recipe passes fp32 fp32 fp32. With an fp32 DiT there is no autocast on either side, so the fp16 CLIP pooled_projections hard-crashes the engine (mat1 and mat2 must have the same dtype) — the text encoders must run fp32 too.
• precision.py: fix a [Do Not Merge] refactor(fsdp)(4/5): per-parameter weight-precision control for FSDP2 #91 regression — wrap_plan looks up gather_dtypes (resolved on clean pre-LoRA FQNs by design) with post-LoRA FQNs, so every --use-lora recipe dies with KeyError: 'base_model.model.transformer_blocks.0' at FSDP wrap. Strip the PEFT prefix at the lookup. Candidate for cherry-pick into [Do Not Merge] refactor(fsdp)(4/5): per-parameter weight-precision control for FSDP2 #91; I will drop it here once it lands there.

Why

• Boundary opt-in: the trajectory serializes latents as fp32 while the SD3 rollout DiT consumed dit-precision; under the passthrough default the train entry dtype diverges from the rollout forward. The policy mirrors what the rollout actually feeds, dump-verified on paired GRPO runs: 2680/2680 tensors bit-exact (cond + uncond CFG branches), 0 dtype mismatches, comparator 2680 passed / 0 failed. Casting the timestep instead rounds the value itself (fp16(857.692) == 857.5); the merged CLIP+T5 embeds enter uncast (T5 runs fp32).
• fp32 recipe: --fsdp-master-dtype already defaults to fp32; this flips the unified knob (train forward/gather + rollout dit) to fp32. With a fp32 train forward the LoRA IPC sync keeps SGLANG_DIFFUSION_LORA_MERGE_FP32=1, closing the last precision gap in the weight-sync path, and the deterministic e2e standard is anchored on fully fp32 numerics.

Validation

• The wrap-plan KeyError reproduced and the fix verified with a minimal CPU repro (bare and PEFT-wrapped module trees) on torch 2.x.
• Paired-dump alignment runs (fp16, LoRA r=32 gaussian, 10-step CFG, SDE candidate step 3, 2×H200): train↔rollout bit-exact end to end with the opt-in policy; entry dtypes hidden_states fp16/fp16, timestep fp32/fp32, encoder_hidden_states fp32/fp32.
• SD3 OCR e2e standard re-record for the fp32 recipe: dispatched via record-e2e-standards.yml on the h200/3gpu runner class (first run 30895256691 hit the wrap-plan KeyError; re-dispatched as 30896211680, which then hit the fp16-CLIP-vs-fp32-DiT engine crash; third dispatch 30897866080 with fp32 text encoders succeeded); the recorded standard is committed. Step-1 train/log_prob_mean_abs_diff = 4.8e-08 (~20x tighter than the fp16-era standard), the parity read before any optimizer step.

Files

• miles/backends/fsdp_utils/configs/sd3.py — SD3 input_dtype_policy opt-in (one comment line + one literal).
• scripts/run-diffusion-grpo-sd3-ocr-sglang.sh — recipe precision default fp16 → fp32.
• miles/backends/fsdp_utils/precision.py — wrap-plan gather-dtype lookup through the PEFT prefix (one line).
• miles/utils/arguments.py + miles/backends/sglang_diffusion_utils/sglang_diffusion_engine.py — text-encoder precision passthrough (2 lines).

Checklist

• pre-commit run --all-files — green in PR CI
• Added/updated tests — none added; the policy mechanism is covered by tests/fast/backends/fsdp_utils/test_input_dtype_policy.py (registered, stage-a-cpu), a config literal needs no extra test
• CPU suites and stage-b/c GPU e2e (incl. SD3 vs the new fp32 standard, bit-for-bit) — all green in PR CI
• python3 train_diffusion.py --help parses (verified on H200 devbox); --sglang-text-encoder-precisions is an existing sglang CLI arg surfaced by the auto-prefix, now forwarded to the engine
• CLI reference docs — n/a, no new flag
• Example walkthrough — n/a, no new example