docs(agentx): correct the warmup-grace CHANGELOG to what shipped

[zengleixin-amd]

Summary

Two sentences in the #1309 CHANGELOG entry describe behaviour that later commits in that same PR removed. The entry was written against the first version of the change and never re-synced, so it now documents a safety property that does not exist.

Documentation only — no code change.

What is wrong

1. The anchor does not default to 8; the scaling is opt-in.

The entry says:

AGENTX_WARMUP_GRACE_CONC declares the concurrency the grace was measured at and defaults to 8, so every existing configuration derives exactly what it derived before.

baseline.py::agentx_warmup_grace_sec (main):

if not _agentx_positive_int(src, "AGENTX_WARMUP_GRACE_CONC"):
    return grace

No anchor → no scaling at all. The stated conclusion happens to still hold, but for the opposite reason. The problem is the mechanism an operator infers from it: read as "CONC scaling is automatic", a grace measured at conc=8 is expected to scale at CONC=32 and does not. That branch also logs nothing, so the round is silently under-budgeted.

Concretely, using this entry's own Kimi-K3 measurements (conc=8 → ~3000s, conc=16 → ~5000s): a 3600s grace at CONC=32 needs roughly 10000s and gets 3600s. The round does not fail — it reports a prefix-reuse figure taken before the cache filled, which is the exact failure #1309 set out to fix.

2. Sweep variants do not re-derive their grace.

The entry says:

A sweep variant re-derives from its own CONC after the variant envs are merged, since the switch runs before that concurrency exists.

That re-derivation was reverted. _grid_runner.py now carries the opposite instruction and the reason:

# NOT re-derived per variant. ...
# Doing that here is WRONG while the two caps above it are session-scaled.
# ...the round is SIGKILLed at 10800s while warmup is still draining --
# strictly worse than leaving the grace alone, which is what shipped before.

Why bother with a CHANGELOG note

Because it is the only human-facing documentation this knob has:

CHANGELOG.md                             <- the only description
docs/                                    <- no AGENTX_* variable appears at all
aiperf_client.sh header                  <- correctly omits it (consumed orchestrator-side)

There is nothing to fall back on, and the wrong direction is the dangerous one — it promises protection that is not there, in the one area where being wrong yields plausible-but-invalid numbers rather than a crash.

Documenting the whole AGENTX_* family in docs/reference/environment-variables.md is a real gap, but a separate and much larger job; adding one knob to a reference that documents none of its siblings would be worse than leaving it. This PR only stops the one description that exists from being wrong.

Scope

• The rewritten paragraph states the opt-in rule, why assuming the anchor was rejected, and what happens when you run above the concurrency you measured at.
• A new Not covered clause records that sweep variants stay session-scaled, and why.

Test plan

• No code changed — git diff --stat is CHANGELOG.md | 38 +++----
• Both claims verified against origin/main source, not against the branch they were written on