fix(cosmos): Use same E2E timeout/availability strategy defaults as reads for PPAF writes

Write hedging should use identical timeout/threshold values as read hedging under PPAF.
The availability strategy behavior should be symmetric — 410s and other transient errors
are handled by the retry pipeline independently of the hedging thresholds.

DIRECT mode (default tcpNetworkRequestTimeout=5s):
- Overall E2E timeout: networkRequestTimeout + 1s = 6s
- Threshold: min(networkRequestTimeout/2, 1s) = 1s
- ThresholdStep: min(threshold/2, 500ms) = 500ms

GATEWAY mode (default httpNetworkRequestTimeout=60s):
- Overall E2E timeout: min(6s, httpNetworkRequestTimeout) = 6s
- Threshold: min(timeout/2, 1s) = 1s
- ThresholdStep: min(threshold/2, 500ms) = 500ms

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: jeet1995

.gitignore

@@ -126,3 +126,4 @@ stress-test-addons*
 
 # Temp typespec files
 TempTypeSpecFiles/
+.dev-tracker/

sdk/cosmos/azure-cosmos/src/main/java/com/azure/cosmos/implementation/RxDocumentClientImpl.java

@@ -7693,40 +7693,8 @@ private CosmosEndToEndOperationLatencyPolicyConfig evaluatePpafEnforcedE2eLatenc
 
     /**
      * Evaluates whether a PPAF-enforced E2E latency policy should be applied for write operations.
-     *
-     * <h3>Timeout Design Rationale (DIRECT mode)</h3>
-     * <p>With default {@code tcpNetworkRequestTimeout} of 5s:
-     * <ul>
-     *   <li><b>Overall E2E timeout = 4x networkRequestTimeout (20s)</b>:
-     *       When a 410 (Gone) is returned, {@code GoneAndRetryWithRetryPolicy} retries with
-     *       exponential backoff (typically 3-4 retries over ~6-8s). The primary Mono consumes
-     *       this budget. The hedge fires at 1x timeout (5s), leaving 15s for the hedged request
-     *       to complete — ample margin for address resolution + transport + response.</li>
-     *   <li><b>Threshold = 1x networkRequestTimeout (5s)</b>:
-     *       Gives the primary write region a full network-timeout window before hedging.
-     *       This avoids premature hedging on transiently slow writes (which are common under
-     *       high throughput) while still firing the hedge well before the E2E timeout.</li>
-     *   <li><b>ThresholdStep = min(threshold/2, 1s)</b>:
-     *       Standard step between consecutive hedge regions (matching read pattern).</li>
-     * </ul>
-     *
-     * <h3>Timeout Design Rationale (GATEWAY mode)</h3>
-     * <p>With default {@code httpNetworkRequestTimeout} of 60s:
-     * <ul>
-     *   <li><b>Overall E2E timeout = 1x httpNetworkRequestTimeout (60s)</b>:
-     *       Gateway mode already bounds requests at the HTTP layer. Using the full timeout
-     *       ensures the E2E timeout doesn't prematurely kill writes that are proceeding
-     *       normally but slowly.</li>
-     *   <li><b>Threshold = min(timeout/2, 2s) = 2s</b>:
-     *       Aggressively hedges after 2s — justified because gateway mode has higher inherent
-     *       latency and the 60s overall budget provides ample room.</li>
-     * </ul>
-     *
-     * <h3>Why separate from reads</h3>
-     * <p>Writes are inherently more expensive (quorum replication, conflict resolution).
-     * The cost of a redundant hedged write is higher than a redundant read. The more generous
-     * threshold (1x vs 0.5x for reads) ensures hedging only activates when the primary
-     * region is genuinely unresponsive, not merely slow under load.</p>
+     * Uses the same timeout/threshold values as the read policy — the availability strategy
+     * hedging behavior should be symmetric for reads and writes under PPAF.
      */
     private CosmosEndToEndOperationLatencyPolicyConfig evaluatePpafEnforcedE2eLatencyPolicyCfgForWrites(
         GlobalPartitionEndpointManagerForPerPartitionAutomaticFailover globalPartitionEndpointManagerForPerPartitionAutomaticFailover,
@@ -7745,14 +7713,9 @@ private CosmosEndToEndOperationLatencyPolicyConfig evaluatePpafEnforcedE2eLatenc
 
                 checkNotNull(networkRequestTimeout, "Argument 'networkRequestTimeout' cannot be null!");
 
-                // For writes, use 4x network timeout as the overall E2E timeout.
-                // This needs to be generous because GoneAndRetryWithRetryPolicy (for 410 errors)
-                // retries with backoff, consuming budget. The hedge fires at 1x network timeout,
-                // leaving 3x for the hedged request to complete after retry overhead.
-                Duration overallE2eLatencyTimeout = networkRequestTimeout.multipliedBy(4);
-                // Threshold = full network timeout — give primary region a fair chance
-                Duration threshold = networkRequestTimeout;
-                Duration thresholdStep = Utils.min(threshold.dividedBy(2), Utils.ONE_SECOND);
+                Duration overallE2eLatencyTimeout = networkRequestTimeout.plus(Utils.ONE_SECOND);
+                Duration threshold = Utils.min(networkRequestTimeout.dividedBy(2), Utils.ONE_SECOND);
+                Duration thresholdStep = Utils.min(threshold.dividedBy(2), Utils.HALF_SECOND);
 
                 return new CosmosEndToEndOperationLatencyPolicyConfigBuilder(overallE2eLatencyTimeout)
                     .availabilityStrategy(new ThresholdBasedAvailabilityStrategy(threshold, thresholdStep))
@@ -7763,11 +7726,10 @@ private CosmosEndToEndOperationLatencyPolicyConfig evaluatePpafEnforcedE2eLatenc
 
                 checkNotNull(httpNetworkRequestTimeout, "Argument 'httpNetworkRequestTimeout' cannot be null!");
 
-                // For gateway writes, use the full HTTP request timeout as the E2E timeout
-                Duration overallE2eLatencyTimeout = httpNetworkRequestTimeout;
+                Duration overallE2eLatencyTimeout = Utils.min(Utils.SIX_SECONDS, httpNetworkRequestTimeout);
 
-                Duration threshold = Utils.min(overallE2eLatencyTimeout.dividedBy(2), Duration.ofSeconds(2));
-                Duration thresholdStep = Utils.min(threshold.dividedBy(2), Utils.ONE_SECOND);
+                Duration threshold = Utils.min(overallE2eLatencyTimeout.dividedBy(2), Utils.ONE_SECOND);
+                Duration thresholdStep = Utils.min(threshold.dividedBy(2), Utils.HALF_SECOND);
 
                 return new CosmosEndToEndOperationLatencyPolicyConfigBuilder(overallE2eLatencyTimeout)
                     .availabilityStrategy(new ThresholdBasedAvailabilityStrategy(threshold, thresholdStep))