[Bugfix][Router] Use supplied timestamp for avg_latency and drop completed request maps

[nisaral]

What

RequestStatsMonitor.on_request_complete recorded time.time() - start instead of timestamp - start, so avg_latency included monitor overhead and was untestable. Completed (engine_url, request_id) entries were also left in request_start_time and first_token_time.

Why

process_request() already captures end_time and passes it in. TTFT already uses the supplied timestamp; latency should too.

Test

src/tests/test_request_stats.py — latency equals the injected interval; maps are empty after complete; missing start is a no-op.

[gemini-code-assist]

Code Review

This pull request addresses a memory leak in RequestStatsMonitor by popping completed requests from the tracking maps and corrects the latency calculation to use the provided timestamp instead of time.time(). It also introduces unit tests for these behaviors. The reviewer suggested updating decoding_length_monitors using the popped first_token_time so that avg_decoding_length is correctly calculated, along with adding corresponding assertions in the tests.

[gemini-code-assist]

The decoding_length_monitors are never updated in the codebase, meaning avg_decoding_length will always return -1. Since first_token_time is popped here, we can use it to calculate the decoding length (timestamp - first_token_time) and update decoding_length_monitors accordingly.

        request_start_time = self.request_start_time.pop(key, None)
        first_token_time = self.first_token_time.pop(key, None)
        if request_start_time is not None:
            self.latency_monitors[engine_url].update(
                timestamp, timestamp - request_start_time
            )
        if first_token_time is not None:
            if engine_url not in self.decoding_length_monitors:
                self.decoding_length_monitors[engine_url] = MovingAverageMonitor(
                    self.sliding_window_size
                )
            self.decoding_length_monitors[engine_url].update(
                timestamp, timestamp - first_token_time
            )

[nisaral]

Taken. avg_decoding_length is documented as time from first token to completion, and popping first_token_time was discarding the only timestamp that can fill it. on_request_complete now updates decoding_length_monitors when a first-token time exists; requests that complete before the first token still leave the gauge at -1.

[gemini-code-assist]

Add an assertion to verify that avg_decoding_length is correctly calculated and updated.

Suggested change

	assert stats[URL].avg_latency == pytest.approx(1.5)
	assert stats[URL].avg_latency == pytest.approx(1.5)
	assert stats[URL].avg_decoding_length == pytest.approx(1.3)

[nisaral]

Added. The existing timestamps give decode duration 101.5 - 100.2 = 1.3, plus a case with no first token that stays at -1.