[Code Health] dep_gen: DepGenRecord carries no per-dep DepFlags, so explicit add_dep_wait edges are mis-recorded in deps.json

[ChaoZheng109]

Category

Observability / DFX accuracy (tech debt)

Component

dep_gen — DepGenRecord capture + dep_gen_replay (tensormap_and_ringbuffer)

Description

PR #1806 (part of #1375) added orthogonal DepFlags { WAIT, RETAIN } per fanin edge and
threaded them into deps.json and the dep_gen differential gate. Creator and tensormap
(modifier) edge flags are re-derivable at replay time (arg_types + owner_task_id +
tensormap), so those are recorded accurately.

Explicit-dep flags are not. DepGenRecord records explicit_deps[] as bare
PTO2TaskId::raw values with no per-dep DepFlags. So the replay hard-codes every explicit
edge to DEP_WAIT | DEP_RETAIN. Consequences:

• A dep added via the ordering-only CoreTaskArgsWithDeps::add_dep_wait() API (a genuine
  ["wait"] edge at runtime) is recorded in deps.json as ["wait","retain"].
• The dual-pass differential gate cannot catch an explicit-kind divergence, because both the
  oracle and the annotated pass read the same hard-coded constant.

This is functionally harmless (runtime behaviour is correct; only the DFX artifact is
imprecise) and is currently documented as a known replay limitation in docs/dfx/dep-gen.md,
but it means deps.json cannot be trusted for the WAIT/RETAIN classification of explicit
edges.

Related, and better handled inside the follow-up reduction PR rather than here: dep_gen
reconstructs edges by re-running compute_task_fanin (construction only), so it does not model
the planned device-side transitive reduction. Once reduction lands, deps.json flags will be
the pre-reduction values unless the replay also models the reduction. (Tracking that with the
reduction work, noted here for context.)

Location

• src/common/platform/include/common/dep_gen.h — struct DepGenRecord (has
  explicit_deps[], no kinds; static_assert(sizeof(DepGenRecord) == 4672) locks the wire
  layout)
• src/{a2a3,a5}/runtime/tensormap_and_ringbuffer/host/dep_gen_replay.cpp — hard-codes
  explicit edges to DEP_WAIT | DEP_RETAIN
• src/common/platform/shared/host/dep_gen_collector.cpp — capture path
• docs/dfx/dep-gen.md — documents the limitation

Proposed Fix

Add explicit_dep_kinds[DEP_GEN_MAX_EXPLICIT_DEPS] (one uint8_t DepFlags per explicit dep)
to DepGenRecord (and the overflow-record chain), populate it at capture, read it in
dep_gen_replay instead of the constant, and make the differential gate explicit-kind-aware.
Update the sizeof(DepGenRecord) static_assert + docs/dfx/dep-gen.md, and drop the
"explicit edges always recorded as wait+retain" caveat. Priority is low — only the DFX
artifact accuracy for add_dep_wait edges is affected, not runtime correctness.

[ChaoZheng109]

@zmnobug