chore(agentx): bump AIPerf to agentx-v1.0.4 [no breaking changes, no perf effect]

[cquil11]

Summary

• bump utils/aiperf from agentx-v1.0.3 (bcfc235c552a17de6d1a7a5d2345cf012401588d) to agentx-v1.0.4 (49634d1c6c2e6dcc18979d704a65e7c7064efc49)
• release: https://github.com/SemiAnalysisAI/aiperf/releases/tag/agentx-v1.0.4
• upstream fix: fix(agentx): accept either latency signal as liveness aiperf#39

Behavior

AgentX keeps the 98% profiling-coverage threshold but now accepts either late TTFT or late ITL as proof of global server activity. This prevents healthy long-running streams from failing solely because no new request starts near the profile boundary. Runs where neither signal reaches the cutoff remain invalid.

This corrects the false-positive validation seen in:

• https://github.com/SemiAnalysisAI/InferenceX/actions/runs/31229823255/job/93031505836
• https://github.com/SemiAnalysisAI/InferenceX/actions/runs/31353823062/job/93349749030

Compatibility and performance

• no breaking changes
• no performance-related changes
• no changes to request payloads, workload pacing, concurrency, latency calculations, routing, or inference-server execution

Validation

• verified agentx-v1.0.4 resolves to 49634d1c6c2e6dcc18979d704a65e7c7064efc49
• verified the release is public, non-draft, and non-prerelease
• upstream: 103 focused unit tests and all applicable pre-commit hooks passed
• verified the InferenceX diff changes only utils/aiperf

[claude]

Claude finished @cquil11's task in 1m 13s —— View job

---

Review of PR #2548

• Gather context (merged submodule-bump PR, 1 file changed)
• Verify the utils/aiperf submodule pointer change
• Post review

LGTM - no blocking issues found.

The diff is exactly one gitlink change in utils/aiperf: bcfc235c552a17de6d1a7a5d2345cf012401588d → 49634d1c6c2e6dcc18979d704a65e7c7064efc49, matching the SHAs and release (agentx-v1.0.4) stated in the PR description. No other files changed, so none of the master-config, perf-changelog, or benchmark-script checks apply. A validation-logic-only change in the profiling-coverage check (accepting late TTFT or late ITL) is consistent with the "no perf effect" claim.

Note: I could not independently re-resolve the upstream tag to its commit SHA from this environment (network gh api/git ls-remote calls require approval), so tag→SHA verification relies on the PR description's stated validation.

[claude]

LGTM — straightforward submodule pointer bump for utils/aiperf (agentx-v1.0.3 → agentx-v1.0.4).

Extended reasoning...

Overview

The PR changes a single line: the gitlink for the utils/aiperf submodule, advancing it from commit bcfc235c5 to 49634d1c6 (tag agentx-v1.0.4). No other files in InferenceX are touched. The upstream change (per the linked release/PR) adjusts AgentX's profiling-coverage validation to accept either late TTFT or late ITL as proof of server activity near the profile boundary, fixing false-positive run invalidations observed in two linked CI runs.

Security risks

None. This is a pointer bump to a vetted internal submodule (SemiAnalysisAI/aiperf) with a tagged, non-draft, non-prerelease release. No new code is introduced into this repo, no changes to request payloads, auth, or execution paths within InferenceX itself.

Level of scrutiny

Low. Submodule bumps of this kind are mechanical — the actual logic change lives and was tested upstream (103 unit tests, pre-commit hooks per the description), and the InferenceX-side diff is a single gitlink line. The PR description explicitly confirms no breaking changes, no perf impact, and no changes to workload pacing/concurrency/routing.

Other factors

No bugs were found by the bug hunting system, and the diff is minimal and self-contained enough that a human reviewer would add little beyond confirming the commit hash resolves correctly, which the author already verified.