feat(session-replay): upload replay envelope on crash with crashpad, breakpad and inproc backends

[tustanivsky]

This PR adds session-replay envelope upload during crash handling for the crashpad (Windows/Linux), breakpad, and inproc backends, extending the native-crash-daemon-only implementation from #1809 to all backends.

The delivery model differs from the daemon: since these backends handle the crash in a dying process (or, for crashpad, hand off to a handler process that only performs multipart minidump uploads), the crash handler consumes the staged <database>/replays/replay-<id>.{mp4,json} files at crash time and persists the resulting replay_video envelope into the run directory via the disk transport. sentry__process_old_runs then uploads it on the next launch. Because the staged replay files are consumed during crash handling, embedders that clear replays/ before sentry_init (e.g. sentry-unreal) remain unaffected.

Key changes

• breakpad / inproc

  • After the crash envelope is prepared in breakpad_backend_callback / process_ucontext_deferred, flush the pending replay through a run-directory disk transport.
  • Source the replay_id from the scope-applied crash event (contexts.replay.replay_id).
  • Execute the flush before dispatching to an external crash reporter so the replay is included on that path as well.

• crashpad (first-chance handler, Windows/Linux)

  • Perform the same run-directory disk transport flush after the session dump.
  • Take an additional reference on the crash event before flush_scope_from_handler, since flush_scope_to_event releases its reference. The scope-applied event is then used to source the replay flush.

• src/sentry_session_replay.h

  • Updated the documentation from "native-daemon-only" to describe both delivery modes:
    • live transport (same-session)
    • run-directory disk transport (next launch)
  • sentry__session_replay_flush_pending itself is unchanged, as it was already transport-agnostic.

• Tests

  • Moved REPLAY_ID, stage_replay(), and is_replay_envelope() into shared test helpers (tests/__init__.py).
  • Factored envelope validation into assert_replay_envelope() (tests/assertions.py).
  • Fixed test_native_replay_orphan_not_flushed, which incorrectly passed an Envelope where raw bytes were expected, making its no-replay assertion vacuously true.

Testing

• Added test_crash_replay_envelope_http (parameterized over inproc and breakpad):

  • Crash with a staged replay.
  • Verify staged replay files are consumed during crash handling.
  • Verify the next launch uploads both the crash envelope and a fully validated replay_video envelope.

• Added test_crashpad_replay_envelope:

  • Verify same-session minidump upload.
  • Verify next-launch replay_video envelope upload.
  • Assert staged replay files are consumed during crash handling.
  • Skipped on macOS (no SetFirstChanceExceptionHandler).

• Existing native-daemon tests (test_native_replay_envelope, test_native_replay_orphan_not_flushed) continue to pass unchanged using the refactored helpers.

• Verified locally on Windows (MSVC).

Related items

• feat(session-replay): upload replay envelope from the native crash daemon #1809
• Add integration with Session Replay product for native platforms sentry-unreal#1445

Known limitations

• macOS Crashpad and Windows WER/fast-fail crashes do not include replay data because no in-process hook executes.

#skip-changelog

[github-actions]

	Messages
📖	Do not forget to update Sentry-docs with your feature once the pull request gets approved.

Generated by 🚫 dangerJS against 7a17df3