PERF: G1Flat training throughput — contact sensor pooled reset + event sync elimination

[kioxli]

Summary

• contact_history: pooled allocations by dtype/shape family (10 separate tensors → 6 pools); reset() fused from 10 CUDA writes → 1 _foreach_zero_ (full reset) / 6 broadcast writes (subset reset)
• event_manager: interval_step == 1 events take an env_ids=None fast path, eliminating a per-control-step GPU nonzero + host len() device synchronization

Why

G1 × 4096 NoRender training is bound by the host-side Python orchestration (~600k calls/control step, GPU busy <25%). The contact sensor reset chain (36 advanced-indexing writes per episode reset, ×26 resets/iteration early in training) and the per-step event scheduling sync were the two largest measurable contributors.

Benchmark

G1 × 4096, NoRender + newton physics, RTX 5090 D, 2 warmup + 10 iterations, passed, compiles=0:

Config	train_e2e_sps
main (7ae9a5012)	80,568
+ this PR	99,398 (+23%)

GPU peak 9.1 GiB / 24.5 GiB.

[greptile-apps]

Confidence Score: 4/5

The PR needs the explicit repository requirements for the new test file and cache-reuse coverage satisfied before merging.

Findings

1.  New test lacks required conventions ▶
2.  Cache reuse remains untested ▶

Fix with agent prompt

### Issue 1
tests/gym/envs/managers/test_event_manager_interval_fastpath.py:1
This new file omits the required DexForce 2021–2026 Apache 2.0 copyright header and `from __future__ import annotations`. The multiline `self.calls.append` at lines 14–16 also needs the repository-required Black formatting. These requirements must be satisfied before merging.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

### Issue 2
tests/gym/envs/managers/test_event_manager_interval_fastpath.py:43
The test calls `apply()` only once, so it would pass if the interval-one path went back to allocating a new ID tensor on every control step. The repository requires focused tests that prove new production behavior. Assert that successive calls reuse the cached tensor to protect this throughput change before merging.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

The PR pools contact-history storage to reduce reset writes and replaces interval-one event ID allocation with a cached all-row tensor.

• Adds a full contact-history reset test and an interval-one ID test.
• The new test file needs repository-required formatting and metadata, and the cache-reuse behavior still needs a focused assertion.

_{Reviews (5) · Last reviewed commit: "FIX: address maintainer review — focused..."}

[greptile-apps]

New test lacks required conventions

This new file omits the required DexForce 2021–2026 Apache 2.0 copyright header and from __future__ import annotations. The multiline self.calls.append at lines 14–16 also needs the repository-required Black formatting. These requirements must be satisfied before merging.

Context Used: CLAUDE.md (source)

Prompt To Fix With AI

This is a comment left during a code review.
Path: tests/gym/envs/managers/test_event_manager_interval_fastpath.py
Line: 1

Comment:
**New test lacks required conventions**

This new file omits the required DexForce 2021–2026 Apache 2.0 copyright header and `from __future__ import annotations`. The multiline `self.calls.append` at lines 14–16 also needs the repository-required Black formatting. These requirements must be satisfied before merging.

**Context Used:** CLAUDE.md ([source](https://github.com/dexforce/embodichain/blob/main/CLAUDE.md))

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[greptile-apps]

Cache reuse remains untested

The test calls apply() only once, so it would pass if the interval-one path went back to allocating a new ID tensor on every control step. The repository requires focused tests that prove new production behavior. Assert that successive calls reuse the cached tensor to protect this throughput change before merging.

Context Used: CLAUDE.md (source)

Prompt To Fix With AI

This is a comment left during a code review.
Path: tests/gym/envs/managers/test_event_manager_interval_fastpath.py
Line: 43

Comment:
**Cache reuse remains untested**

The test calls `apply()` only once, so it would pass if the interval-one path went back to allocating a new ID tensor on every control step. The repository requires focused tests that prove new production behavior. Assert that successive calls reuse the cached tensor to protect this throughput change before merging.

**Context Used:** CLAUDE.md ([source](https://github.com/dexforce/embodichain/blob/main/CLAUDE.md))

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!