feat(dsa): reuse full-layer top-k on shared GLM-5.2 decode layers

[Andrewxu313]

What

Wire GLM-5.2 DSA indexer top-k reuse into the eager (non-CUDA-graph) decode
selector. In build_glm5_dsa_flashmla_inputs, skip the indexer-K compute and
aux-KV update on shared layers (indexer is None), reuse
type(wrapper)._dsa_prev_topk_indices when skip_topk, and publish it from
full layers when next_skip_topk. index_topk is hoisted to a local so the
sparse gather stays None-safe on shared layers.

Why

GLM-5.2 runs its DSA indexer on the full layers only; the shared layers carry
no indexer weights and reuse the previous full layer's top-k. This is the
selector-side half of that reuse. GLM-5 has no such schedule
(dsa_layer_skips_topk is False for every layer), so every new branch is dead
for it and its behavior is unchanged. The CUDA-graph segment builders still
dereference the indexer unconditionally and are handled separately. Stacked on
the model-side hooks in #310 — review/merge that first.

Type of Change

• model
• kernel
• core
• fix
• infra
• docs

File changes

File	Δ	Note
batchgen/attention/dsa/glm5_decode_selector.py	mod	skip indexer-K on shared layers; reuse/publish carried top-k; None-safe sparse gather

Checklist

• Every changed file traces to this task; the diff is surgical and one concern (§3).
• Changed files stay within the declared type's allowlist — a model/kernel PR does not touch the scheduling/scaffolding layer (§2.5–§2.6).
• The File changes table lists every changed file and matches the diff (§2.5).
• No debug_*/scratch_*/tmp_* scripts, BATCHGEN_* env-guards, stray print(), committed artifacts, or Co-Authored-By (§1/§4). (bench_* benchmarks are fine.)
• Tests added/updated under tests/; touched modules' MODULE.md updated if the public API changed (§2).
• Ran bash .github/workflows/scripts/check-pr-hygiene.sh locally (against the tairanxu/glm52-model base); the CI hygiene check is green.

🤖 Generated with Claude Code