Write solution(metrics, k, error_threshold, latency_threshold). Given monitoring samples [requests, errors, p95_latency] (one per minute, chronological), a window size k, error-rate threshold, and latency threshold, return the starting indices of all contiguous windows of exactly k samples that breach the SLA.

For a window: total_requests/total_errors = sums over window; max_latency = max p95_latency in window. error_rate = total_errors/total_requests (0 if total_requests = 0). Window breaches iff error_rate > error_threshold OR max_latency > latency_threshold (both strict; exact ties are NOT breaches).

If length(metrics) < k, return empty list.

Constraints: 0 ≤ length(metrics) ≤ 200,000; 1 ≤ k ≤ 200,000; requests/errors/p95_latency integers with 0 ≤ requests ≤ 1,000,000, 0 ≤ errors ≤ requests, 0 ≤ p95_latency ≤ 1,000,000; thresholds ≥ 0 with ≤4 decimal digits. Must run in ~O(length(metrics)) — an O(length × k) solution will TLE on hidden large inputs. Compare using genuine double division total_errors / total_requests > error_threshold, not the rearranged multiplication form (which can misjudge exact ties due to floating-point rounding).

Example 1

metrics = [[100,40,100],[100,0,100],[100,0,100],[100,0,100],[100,0,400],[100,0,100]]
k = 3, error_threshold = 0.05, latency_threshold = 300

Output: [0, 2, 3]

Example 2

metrics = [[0,0,100],[0,0,100],[4,1,500],[4,1,500]]
k = 2, error_threshold = 0.25, latency_threshold = 500

Output: [] (both computed values are exactly equal to thresholds — not breaches)

Example 3

metrics = [[100,10,100],[100,10,100]]
k = 2, error_threshold = 0.05, latency_threshold = 500

Output: [0]