[Refactor] Record what a bound stage path does instead of reading it back

[ch-wan]

This PR is part of a stack (oldest at bottom):

• [Refactor] Declare every model's stages from a shared function and drop the meta-device neighbour build #43460
• [Refactor] Declare pipeline neighbours from a shared declaration function #43459
• [Refactor] Decide an exit's fixed facts when the stage is built #43458
• [Refactor] Bind a layer stack in one pass over each stage's data flow #43457
• [Refactor] Decide input-scattered eligibility and stage rejections in one place #43456
• [Refactor] Make residual read and update capabilities explicit and checked #43455
• [Fix] Keep residual rows and sums consistent across appends and pipeline handoffs #43454
• -> [Refactor] Record what a bound stage path does instead of reading it back #43453

Motivation

Several checks of a bound stage boundary path (construction checks, the residual-preservation certificate, auxiliary capture) found out what the path does by inspecting the bound partials: partial.func is X and their keywords. Wrapping a step, or binding it differently, silently changes what those checks see. The arriving capability was also carried as a tuple with a per-update dispatcher, although every arrival has exactly one.

Modifications

• Bind a single arriving capability on stage boundaries. EdgeContract.arriving_plain_add is a single value; the capability dispatcher, the multi-capability branch of bind_entry, and the tests that only exercised them are removed, together with the unused fetch_hidden_states path on the attention inputs.
• Record what a bound stage boundary path does instead of reading it back. Entry selection returns what it already knows: whether the input move gathers over attention TP with the stage's own gather (input_gather_declared), whether a capture may keep that move's output without copying (input_retainable), and which fused kernels the step tries (ffn_fusions); the exit move records its attention-TP gather (StagePath.output_gathers_attn_tp). The checks read these fields.

No bound path or capture result changes.

Accuracy Tests

How the stack is checked:

• Unit tests: the 39 layer-boundary test files pass at the head of every PR in the stack, apart from the two AMD-only test_aiter_allreduce_fusion_amd cases that fail on NVIDIA on main too; the test files that changed most recently pass on every commit. The full test/registered/unit suite (926 files, one process per file) shows no failure on the stack's head that main does not have.
• Plan comparison: an out-of-tree harness binds every stage plan for seven stack shapes (decoder layers with and without an output transform, MHC, Nemotron-H patterns, Kimi K3, Qwen4-exp, NemotronNAS-style attention-only layers, single-stage appends) over 23 parallel configurations, four boundary reductions and one or two pipeline ranks (3,928 scenarios), and compares each bound plan leaf by leaf with the parent commit's.
• GPU: dummy weights are seeded per TP rank and scaled (0.05, norms 1.0), so that a sum taken twice or a mixture of rank partials changes the output; the default dummy weights give every rank the same shard and would hide both.
• GPU results below were measured on these commits as based on main 516fd1c1b5 unless marked as on the merge candidate; rebasing onto current main had no conflicts, and the CPU checks were run again after it.
• On the merge candidate (head of this stack against main 961bcf481d, H200, asymmetric dummy weights unless noted): DeepSeek-V2-Lite input-scattered PP2 is bit-identical to PP1 (main crashes on it); Nemotron-H cut between a mixer and its FFN differs from PP1 only at the rounding level of the FFN-to-mixer cuts (main: 0.93 and 2.21); Llama, Qwen3-MoE, GPT-OSS and DeepSeek-V2-Lite are bit-identical to main at PP1 and PP2; GSM8K (real DeepSeek-V2-Lite weights, 250 questions, TP4 with attention DP2 and the dense FFN fully data-parallel) gives 0.360 at PP2, 0.364 at PP1 and 0.368 on main at PP1, medians of three runs whose run-to-run spread is 0.012–0.016. The last revision of the stack changed only comments, docstrings, an equivalent condition and explicitly set attributes after these runs; its bound plans are identical.

For this PR: the plan comparison against main shows no difference in any scenario, and over the whole configuration matrix the new fields equal what the checks used to infer from the partials.

Speed Tests and Profiling

No kernel, collective or forward-path change.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): 🚫 Run #38080721135
Latest PR Test (Extra): 🚫 Run #38080720881
Latest PR Test (AMD ROCm 10): ❌ Run #38080721057