feat(perf-monitor): aggregate CPU/RAM across descendant processes

perf-monitor/README.md

@@ -77,11 +77,39 @@ Only the fields you specify are overridden; others inherit from `[thresholds]`.
 
 ### Metrics
 
-| Metric           | Description                                                |
-| ---------------- | ---------------------------------------------------------- |
-| CPU Avg / P95    | Process CPU usage normalized by logical CPU count (0-100%) |
-| Memory Avg / P95 | Resident Set Size (RSS) in MB                              |
-| Memory Growth    | Last sample minus first sample (detects obvious leaks)     |
+CPU and memory are aggregated across the target process **and all of its
+descendants** (children, grandchildren, ...). This matters for apps that
+spawn helper / renderer processes (e.g. Tauri's WebView2 host on Windows,
+Electron-style multi-process layouts) — measuring only the parent PID
+would miss the bulk of the actual resource usage.
+
+| Metric           | Description                                                              |
+| ---------------- | ------------------------------------------------------------------------ |
+| CPU Avg / P95    | Aggregated CPU usage normalized by logical CPU count (0-100%)            |
+| Memory Avg / P95 | Aggregated Resident Set Size (RSS) in MB across the process tree         |
+| Memory Growth    | Last sample minus first sample (detects obvious leaks)                   |
+
+### Report sections
+
+The text report (and the matching JSON output) is structured to make
+CI logs easy to scan:
+
+1. **Aggregated (process tree)** — avg / P95 / max across the parent +
+   descendants, with per-check `PASS` / `FAIL` against the configured
+   thresholds.
+2. **Per-Process Breakdown** — one block per observed PID showing the
+   process name, how many sample ticks it appeared in, and its own
+   avg / max / P95 for CPU and RSS. The root process is listed first and
+   marked `(root)`; descendants follow in descending order of average CPU.
+3. **Failure Reasons** — printed only when at least one threshold was
+   exceeded; enumerates exactly which checks failed and the offending
+   value vs. the threshold. The JSON output exposes the same list as a
+   `failure_reasons` array.
+
+If the measurement aborts before any samples are collected (binary
+crashed during warmup, etc.), the failure reason and the captured
+child stderr are printed under a `=== Performance Test Failed ===`
+banner so the cause is visible at the bottom of the CI log.
 
 ### Exit Code
 

perf-monitor/src/main.rs

@@ -72,7 +72,10 @@ fn main() -> ExitCode {
       }
     }
     Err(e) => {
-      eprintln!("Error: monitoring failed: {e}");
+      eprintln!();
+      eprintln!("=== Performance Test Failed ===");
+      eprintln!("Failure reason: monitoring aborted before metrics could be collected");
+      eprintln!("Details: {e}");
       ExitCode::FAILURE
     }
   }