feat: add pod and serverless worker log readers

[muhibwqr]

Closes #400.

The REST v2 API already serves pod and worker logs as SSE streams (GET /v2/pods/{id}/logs, GET /v2/serverless/{id}/workers/{workerId}/logs), and runpodctl pod logs wraps them. This adds the Python SDK side. It follows the bounded-snapshot shape agreed in the issue thread, plus a follow mode.

API

# snapshot: backfill + up to max_wait seconds of live output
logs = runpod.get_pod_logs(pod_id, tail=200, source="container", max_wait=5)
# [{"id": "...", "ts": "2026-06-01T12:02:03Z", "source": "container", "line": "Model loaded."}, ...]

# follow: yields as lines arrive; resumes via Last-Event-ID
for entry in runpod.iter_pod_logs(pod_id, tail=0):
    print(entry["line"])

# serverless workers
workers = runpod.get_endpoint_workers(endpoint_id)
logs = runpod.get_endpoint_worker_logs(endpoint_id, workers[0]["id"])

Addition	Notes
get_pod_logs, get_endpoint_worker_logs	Return a list, oldest first. Stop after max_wait seconds, or once the stream has been idle that long. Past max_bytes (default 4 MiB) the oldest lines are dropped, so the newest output (usually the crash) is kept.
iter_pod_logs, iter_endpoint_worker_logs	Generators. When the stream closes or goes idle they reconnect with Last-Event-ID, so lines are neither repeated nor skipped. On reconnect they retry network errors and 429s (honoring Retry-After). Errors on the first connection, and non-429 HTTP errors, raise. Optional max_wait caps total time.
get_endpoint_workers	GET /v2/serverless/{id}/workers. Worker log readers need worker IDs, and the SDK had no way to list them.
QueryError.retry_after	Seconds from Retry-After on any REST error, or None.
rest.read_event_stream	Streaming SSE transport. run_rest_request is unchanged.

Arguments are validated before any request is made (tail 0–5000, source in container/system/None, since as an RFC3339 string or a timezone-aware datetime). This applies to the generators too.

Tests

• Unit tests for the SSE parser (frames split across chunks, CRLF, comments, multi-line data, UTF-8 split mid-character, an incomplete trailing event dropped), the transport, and reconnect/retry logic.
• tests/test_api/test_logs_server.py runs the readers against a local fake of the log endpoints over real sockets, so the requests/urllib3 chunked-read and read-timeout behavior is exercised, not mocked. It covers the idle and busy-stream deadlines, a cut frame, the byte cap, 404, Last-Event-ID resume without gaps, a 429 on reconnect, max_wait, and the worker listing plus worker logs. It takes about 4s.
• Full suite: 839 passed, coverage 94.32%. ruff and pylint report nothing on the changed modules.

Not yet run against the live API. The wire format follows the published API reference and the runpod-mcp stream-pod-logs reader.

Docs: README Logs section, and retry_after in docs/api/handling_errors.md.

🤖 Generated with Claude Code

[muhibwqr]

hi my name is muhib, i am a second year waterloo math student, i thought it would be pretty cool to push a pr here as I wanted to a) find an internship for the winter b) actually try and understand how runpod works