Fix KubernetesPodOperator emitting orphan timestamps for empty container writes

[kaxil]

When a container running under KubernetesPodOperator writes an empty line, kubelet streams it back (with timestamps=True) as "<rfc3339-ts> \n" -- a timestamp followed by a separator space and an empty message. parse_log_line in pod_manager.py called line.strip().partition(" "), which removed the trailing separator space before partitioning, so the function returned timestamp=None and the caller treated the line as a continuation of the previous buffered log record. The bare RFC3339 string was then appended onto the previous message and emitted as a multi-line log where only the first line carried the Airflow [ts] {pod_manager.py:N} INFO - prefix, leaving unprefixed timestamp rows interleaved in task logs:

[2026-05-28T13:07:50.160+0000] {pod_manager.py:520} INFO - [base] first test line
2026-05-28T13:07:57.030578889Z 
2026-05-28T13:07:57.030581518Z
2026-05-28T13:07:57.030642740Z 
[2026-05-28T13:07:57.034+0000] {pod_manager.py:520} INFO - [base] last test line

Downstream that breaks airflow.utils.log.file_task_handler._parse_timestamp, which feeds the line to pendulum.parse after stripping []: malformed fragments from these orphan rows can raise ValueError: month must be in 1..12 and fail the task entirely.

Closes #36571.

Root cause and history

Regressed in #33675 (merged 2023-08-24, shipped in cncf-kubernetes 7.5.0) which replaced the original line.find(\" \") split with a line.strip().partition(\" \") pattern under the banner of a refactor:

- split_at = line.find(\" \")
- if split_at == -1:
-     ...
- timestamp = line[:split_at]
- message = line[split_at + 1 :].rstrip()
+ timestamp, sep, message = line.strip().partition(\" \")
+ if not sep:
+     ...

The pre-refactor implementation correctly handled <ts> \n because find(\" \") matched the separator space directly and the message-side .rstrip() produced an empty string. The new code strips the separator off before partitioning, so the function loses its only signal that the line is well-formed.

This matches the regression window the original reporter described in #36571: the bug appeared after upgrading cncf-kubernetes from 7.4.2 (pre-refactor) to 7.12.0+ (post-refactor) and is still reproducible on current main (10.17.x).

Fix

• parse_log_line no longer pre-strips the line; it rstrip(\"\\n\") only and partitions on the original separator, so empty container writes are recognised as (timestamp, \"\") rather than as continuations. If the partition yields no separator the whole line is tried as a bare timestamp (some kubelet versions emit <ts>\\n with no trailing space), and parse failures fall through to the original return-the-raw-line path. It also catches ValueError, not just ParserError, so a malformed timestamp can never escape into Airflow's downstream parsers.
• The sync (PodManager.fetch_container_logs.consume_logs) and async (AsyncPodManager.fetch_container_logs_before_current_sec) log consumer loops skip emit for empty messages -- the resume marker still advances in the sync path so reconnect-since-time stays correct, but no noisy [base] row is written.

Tests

• Parametrized test_parse_log_line_handles_empty_container_writes covers <ts> \\n, <ts>\\n, and <ts> (no newline). Verified RED on main, GREEN with the fix.
• End-to-end test_empty_container_lines_do_not_pollute_previous_message drives fetch_container_logs with the exact log sequence from the issue and asserts no orphan timestamps land in caplog. Also RED on main, GREEN with the fix.

Gotchas

• Truly empty container output (just \\n) is no longer surfaced as a [base] row. That output carries no information for the task log reader and was previously the trigger for downstream pendulum failures, so dropping it is a net improvement; if a future use case needs to count blank container lines, that's separable work.