Bugzilla

endSession() recorded newtab.closed, drained the shared event buffer and
submitted the newtab ping before it checked whether the session had ever
been visible. The never-shown guard sat about 35 lines further down, so
it only ever suppressed the session_duration bookkeeping: hidden preloaded
newtabs still produced a full newtab_session_end ping carrying a
newtab.closed event for a visit the user never saw.

Observed steady state was three telemetry sessions for one visible newtab,
two of them with is_preloaded: true and no visibility_event_rcvd_ts.
Opening a browser window and closing it submits a newtab_session_end ping
for the never-shown id.

Hoist the guard above the recording so a session with no
visibility_event_rcvd_ts is dropped without emitting anything, and dedent
the timestamp rounding that used to live in the if branch.

The buffer half matters as much as the event: because #eventBuffer is
shared across sessions, these phantom preload session-ends were also the
main thing draining it mid-session, which is why impressions and section
impressions were being attributed to a ping describing a newtab that was
never displayed. The preceding patch adds the shutdown flush that keeps
those events from being lost now that preload teardown no longer sends
them.

test_endSession_no_ping_on_no_visibility_event already described this
behaviour but only asserted over legacy telemetry events, so it passed
throughout. It now also asserts that no newtab ping is submitted and that
newtab.closed is not recorded.

Claude-Session: https://claude.ai/code/session_015UDkPBx7WgtQ8EZ7LDV1X5

endSession() recorded newtab.closed, drained the shared event buffer and
submitted the newtab ping before it checked whether the session had ever
been visible. The never-shown guard sat about 35 lines further down, so
it only ever suppressed the session_duration bookkeeping: hidden preloaded
newtabs still produced a full newtab_session_end ping carrying a
newtab.closed event for a visit the user never saw.