Add: dep_gen deps.json v2 — tensor-annotated edges + differential replay

[ChaoWao]

Summary

Resolves #666 — deps.json now carries per-edge tensor metadata (offset/shape/dtype) and per-task input/output slot info, instead of just task→task IDs. Zero runtime changes; all work lives in the host-side replay + downstream tools.

v2 schema (replaces v1, no fallback):

• tasks[] — per task: task_id, scope, args[] with {idx, type, tensor_id, dtype, shape, offset} per slot. OUTPUT slots omit tensor info (zero blob at submit time).
• tensors[] — one entry per unique (buffer_addr, version); stable FNV-1a tensor_id; raw_shapes for the underlying buffer.
• edges[] — {pred, succ, arg, source} plus tensor_id / consumer_shape / consumer_offset (non-explicit) and producer_shape / producer_offset (source=tensormap).

Self-checking replay — no shotgun-modification risk on compute_task_fanin:

• Per record, two parallel PTO2TensorMap instances run in lockstep.
• Oracle pass drives the canonical compute_task_fanin template (unchanged, zero runtime touch) and collects the producer-id set the runtime would have wired.
• Annotated pass inlines the same STEP A / STEP B logic against a second map but the callback fires with the full PTO2TensorMapEntry& + consumer Tensor* + arg index, capturing per-edge metadata.
• After both passes, the producer-id sets are compared. Divergence → LOG_ERROR + return non-zero, deps.json not written. Anyone who later changes compute_task_fanin semantics will trip the gate immediately and know to mirror the change in the annotated pass.
• INOUT+COVERED remove_entry is mirrored so both maps stay bit-equivalent for the next record.

Viewer (deps_to_graph.py --show-tensor-info):

Replaces per-edge labels with HTML-table task nodes — input rows (blue) on top, identity header in the middle (core-type colored), output rows (orange) on the bottom. Each arg cell is a 4-line block: arg<i> <TYPE> <Tname>:<dtype> + raw: / shape: / offset:. Edges route pred:out_<idx>:e → succ:in_<arg>:w by matching tensor_id. OUTPUT slots backfill their tensor_id from downstream creator edges when unambiguous (marked ? in the row).

Test plan

• pip install --no-build-isolation -e . builds clean on macOS arm64
• python test_dep_gen.py -p a2a3sim --enable-dep-gen (which auto-adds --enable-l2-swimlane) PASSED on a2a3sim — 5 tasks, 7 tensors, 6 annotated edges, fanout ⊆ deps gate green
• Differential check did not fire (oracle ≡ annot for the vector_example workload)
• deps_to_graph.py smoke-tested in both plain and --show-tensor-info modes
• swimlane_converter.load_deps_json correctly projects v2 → {pred → [succ]}
• Hardware run on a2a3 (CI)
• Verify on a larger workload (paged_attention) that overlap=other / partial-tensormap edges still pass the differential check

Issue: #666

[gemini-code-assist]

Code Review

This pull request upgrades the dependency graph generation to version 2, introducing tensor-level annotations such as shapes, offsets, and dtypes. The core replay logic now employs a dual-pass differential check between a canonical oracle and an annotated mirror to guarantee correctness. Corresponding updates were made to the visualization tools, including a new --show-tensor-info mode, and the test suite. Feedback suggests serializing 64-bit integers as strings in the JSON output to ensure interoperability with JavaScript parsers.