feat: record launch latency in statistics for start-only jobs

StartOnly jobs previously forced DurationMS to 0 because GoSentry does
not wait for the process to exit, leaving the Statistics line stuck at
"last 0 ms, avg 0 ms". The runner already measures launch latency (time
to spawn the process) for the History detail; this now returns that
value as the run duration so the existing duration-driven stats pipeline
folds it into last/avg/max.

Sub-millisecond launches still round to 0 and are excluded from the
average, matching prior behavior.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

docs/ARCHITECTURE.md

@@ -89,8 +89,9 @@ flowchart LR
    `runner.RunJob` builds the platform-specific invocation, executes the
    command through the platform shell, captures stdout and stderr, writes one
    timestamped `.log` file, and returns a `domain.RunRecord` containing
-   `DurationMS` (wall-clock milliseconds from start to finish; 0 for
-   `StartOnly` fire-and-forget jobs).
+   `DurationMS` (wall-clock milliseconds from start to finish; for `StartOnly`
+   fire-and-forget jobs it measures launch latency — the time to spawn the
+   process — since there is no exit to wait for).
 
 6. History update:
    When a run goroutine completes, `Service` updates the job's runtime
@@ -132,7 +133,7 @@ in flight increments `JobRuntime.PendingRuns`. When the current run finishes,
 |-------|---------|
 | `RunCount` | total runs recorded |
 | `FailCount` | runs that exited non-zero |
-| `LastDurationMS` | wall-clock time of the most recent run |
+| `LastDurationMS` | wall-clock time of the most recent run (launch latency for `StartOnly`) |
 | `AvgDurationMS` | mean over all runs with a recorded duration |
 | `MaxDurationMS` | longest recorded run |
 

docs/CHANGELOG.md

@@ -2,6 +2,14 @@
 
 All notable GoSentry changes are recorded in this file.
 
+## Unreleased
+
+**Statistics:**
+- `StartOnly` jobs now record launch latency (time to spawn the process) as the
+  run duration instead of a hard-coded `0`, so the Statistics line shows a real
+  last/avg/max for fire-and-forget jobs. Sub-millisecond launches still round to
+  0 and are excluded from the average, as before.
+
 ## 0.11.3 - 2026-06-29
 
 **Reliability fixes from an internal code review: safer runs, a real overlap