ci(cpp): widen memory_per_step_growth_kb threshold to 75% (#874)

## Summary

Bump the regression threshold for `memory_per_step_growth_kb` from the
generic 15% to 75%, so legitimate single-digit-KB feature growth doesn't
false-fail the C++ Benchmarks gate. This is the only metric at a
single-digit-KB scale, where ordinary feature work produces large
*percent* swings on small *absolute* numbers.

## Root cause

`memory_per_step_growth_kb` measures KB of RSS growth per agent loop
step. Pre-VLM main was ~7.2 KB/step. PR #858 (VLM image support, merged
2026-04-24) added the `Image` type, content-block JSON parsing, and new
per-message storage — bumping the metric to ~11.4 KB/step. That's +4 KB
absolute, a perfectly reasonable cost for a major new feature, but it
reads as +58% on the percent scale, well past the 15% generic threshold.

The cached baseline never refreshed because the workflow only saves a
new baseline on a successful main push (`if: github.ref ==
'refs/heads/main' && github.event_name == 'push'`), and every main push
since #858 has either failed (this same gate) or been cancelled. So
**every PR that touches C++ keeps inheriting the stale pre-VLM
baseline** and failing on this one metric.

## Why a threshold change, not a baseline bump

A baseline bump alone is a one-shot fix — the next feature that adds
another KB-level allocation per message will hit the same wall. A wider
threshold for *just this metric* is the structural fix: it acknowledges
that on a metric whose absolute values are tiny, percent swings are
noisy. Other metrics (binary size, latency, peak memory in MB) keep
their tighter thresholds.

## Verification

Most recent failing run shows the new value within the new band:

```
loop_latency_median_us           1026.0   918.0  -10.5%   15%   IMPROVED
memory_baseline_kb               9272.0  9332.0   +0.6%   15%   OK
memory_peak_kb                   9416.0  9560.0   +1.5%   15%   OK
memory_per_step_growth_kb           7.2    11.4  +58.3%   15%   FAIL  <-- this one
```

With the 75% threshold the FAIL becomes OK; everything else is
unchanged.

## Test plan

- [ ] CI green on this PR (the same Windows benchmark job that's been
red on every main push since 2026-04-24)
- [ ] After merge, baseline auto-refreshes on the next main push, future
PRs no longer inherit the stale 7.2 KB baseline

Author: kovtcharov

cpp/benchmarks/bench_utils.h

@@ -180,6 +180,13 @@ inline double thresholdForMetric(const std::string& name) {
     if (name.find("binary_size") != std::string::npos) {
         return 10.0;
     }
+    // Per-step growth is a single-digit-KB metric, so a single feature addition
+    // (e.g. VLM image support adding ~4 KB/step for new content-block storage)
+    // reads as a 50%+ swing. Use a wider band so legitimate feature additions
+    // don't fail the gate while still catching true 2x-style regressions.
+    if (name == "memory_per_step_growth_kb") {
+        return 75.0;
+    }
     // All other metrics: 15% threshold
     return 15.0;
 }