logreader: read segments in order, not discovery order

[lgoyal6]

Description

auto_source accumulates valid_files: dict[int, str] across two nested loops: filenames (rlog, then qlog) at logreader.py:169 and sources at :170. Each pass prunes needed_seg_idxs down to the segments still missing (:178), so a later pass only ever adds the gaps, and those entries land at the end of the dict. The return handed back .values(), which is insertion order, not segment order:

valid_files |= files                                                         # :175
needed_seg_idxs = [idx for idx in needed_seg_idxs if idx not in valid_files]  # :178
if len(needed_seg_idxs) == 0:
  return list(valid_files.values())                                          # :182

Whenever a lower-numbered segment is resolved after a higher-numbered one, the returned list is out of order. Two ways that happens in normal use:

1. AUTO mode falling back to qlogs. If a middle segment's rlog is missing, the rlog pass returns the segments around it and the qlog pass fills the hole last. This is the path that already logs "N/M rlogs were not found, falling back to qlogs for those segments..." (:192), so it is an expected case, not an exotic one.
2. Plain RLOG mode, no fallback involved. The default source list is four sources (:249) and each returns only the segments it has. If the first source is missing an early segment that a later source supplies, the order is wrong.

LogReader.__iter__ (:268) reads the list in the order given, so segments come out shuffled and logMonoTime moves backwards across segment boundaries. sort_by_time does not cover this: it sorts self._ents within a single file (:120).

The existing tests cannot observe it. local_source and local_auto_source return the same path for every segment via dict.fromkeys (test_logreader.py:84, :89), so ordering is invisible, and local_auto_source returns {} for rlogs, which makes the qlog pass fill every segment in one call and therefore in order. test_sort_by_time (:254) reads a single file.

Fix: return the files ordered by segment index. seg_idxs is always ascending, since a descending slice always evaluates to empty (10:0:-1 and 0:6:-2 both give []), so this is exactly the requested order and can never reorder a valid request.

Verification

Three segments where segment 1 has no rlog, resolved through a fake source, so no network or device is needed:

before: ['seg0/rlog.zst', 'seg2/rlog.zst', 'seg1/qlog.zst']
after:  ['seg0/rlog.zst', 'seg1/qlog.zst', 'seg2/rlog.zst']

End to end, same setup with real logs written by save_log and read back through LogReader(..., sort_by_time=True):

before: logMonoTime [0, 2000000000, 1000000000]   # time goes backwards 1s
after:  logMonoTime [0, 1000000000, 2000000000]

Also reproduced with no qlog fallback at all: two rlog-only sources, the first holding segments 1 and 2 and the second holding segment 0, returned [seg1, seg2, seg0] before and [seg0, seg1, seg2] after.

openpilot/tools/lib/tests/test_logreader.py: 56 tests pass (22 skipped, network). ruff, ty, and codespell clean.

[github-actions]

Process replay diff report

Replays driving segments through this PR and compares the behavior to master.
Please review any changes carefully to ensure they are expected.

✅ 0 changed, 66 passed, 0 errors

[lgoyal6]

The two red checks are the fork-runner GL gap, not this change.

Because this PR comes from a fork, the runner selection falls through to GitHub's hosted image (Image: ubuntu-24.04) rather than the namespace-profile machines. That image has no libGLESv2.so.2, so raylib's headless backend cannot load:

File ".../site-packages/raylib/__init__.py", line 33, in _load_cffi
ImportError: libGLESv2.so.2: cannot open shared object file: No such file or directory
ImportError: failed to load raylib headless backend extension _raylib_cffi_headless

Create UI Report dies on the same import. tools/setup_dependencies.sh does not install any GL/GLES packages, so the library has to come from the runner image.

unit tests is otherwise clean, with that one UI test as the only failure:

856 passed, 43 skipped, 1 xfailed, 1 failed in 85.74s
FAILED openpilot.selfdrive.ui.tests.test_raylib_ui.TestRaylibUi.test_raylib_ui

I also ran this workflow against untouched master on my fork as a baseline, and it returns the identical 856 passed, 43 skipped, 1 xfailed, 1 failed with the same single failure, so this change adds none.

Everything that does exercise the change passes: static analysis, build release, build macOS, and process replay (0 changed, 66 passed, 0 errors).