fix(engine): make ragged transport padding objective-safe

[EazyReal]

Problem

After semantic compaction, token-aware splitting can give one data-parallel rank three real microbatches and another only one. FSDP collectives and Megatron/Archon pipeline schedules still require every rank and stage to execute the same number of steps; a shorter rank cannot simply stop without hanging its peers.

Why transport-only padding

Collection should first compact and backfill real work (#1563), but that cannot guarantee equal physical microbatch counts. Truncating to the smallest rank drops valid samples, and changing the loss denominator does not satisfy distributed collectives. Duplicating a real sample changes the objective, zero-length model inputs are not generally valid, and loss_mask=0 is insufficient for DPO, reward-model, or custom objectives whose weights come from sequence structure.

The smallest shared backend mechanism is therefore a model-valid transport row with an explicit marker: it participates in forward/backward and communication, but the engine bypasses objective callbacks and assigns it zero loss weight.

Solution

• add transport-only microbatches only when synchronized execution requires them
• bypass arbitrary objective weight and loss callbacks for transport rows
• keep transport rows out of every loss denominator, metric, normalization population, and returned output
• strip the marker before model execution and support ordinary and tree-packed paths across FSDP, Megatron, and Archon

Scope boundary

#1563 owns semantic membership, compaction, group boundaries, and normalization. This PR acts only after those decisions, at the engine scheduling boundary; transport rows never become rollout members.

Review stack

• Focused nine-file review mirror: EazyReal/AReaL#7
• Dependent upstream PR: fix(rollout): train safely on incomplete groups #1563

GitHub cannot base an upstream PR on a contributor-fork branch, so #1563 remains temporarily cumulative against main and shrinks to its core diff after this foundation lands.

Part of #1559.

Verification

• pytest tests/test_utils.py tests/test_tree_transport.py tests/test_eval_dispatch.py -q — 86 passed
• pre-commit run --all-files
• upstream hosted pre-commit and Ubuntu, macOS, SGLang-extra, and vLLM-extra installation checks passed; Docker was skipped by workflow policy
• independent exact-range transport, Megatron, and Archon review of 08ab30fc..b0dbd4c4 — no actionable findings

A real multi-GPU Megatron/Archon pipeline canary was not available on this host; this is static, CPU/distributed-test, and hosted-install readiness rather than hardware certification.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[EazyReal]

This makes transport-only padding rows safe for arbitrary objectives: they run through forward/backward for cross-rank and pipeline step parity, but are excluded from loss weight, metrics, and outputs. None of this exists on main yet (no TRANSPORT_DUMMY_KEY or allow_transport_padding), and the closest existing path — tree dummy-trie rows — still reaches loss_weight_fn before its skip guard. It's mergeable, tests are included and passing, and CI is green — @sitabulaixizawaluduo, could you take a look?