GeezerVison

Every request the fda client sends now retries transient failures with exponential backoff, controlled by the global --retries flag (FELDERA_RETRIES, default 3, 0 disables). This addresses the fda half of feldera/cloud#1907, where a single connect timeout to a freshly rescheduled pipeline pod failed fda clock-advance and with it the integration suite.

Three commits:

1. Remove the bench result-upload path. fda bench still runs benchmarks and prints results; the upload to benchmarks.feldera.io was unused. Deletes the generated Bencher client, bench_openapi.json, fda's build.rs, and the --upload/BENCHER_* flags.
2. Retry transient request failures. The retry layer lives in feldera-rest-api as a ClientHooks::exec override (progenitor 0.7 → 0.11.2, the newest release on reqwest 0.12), so all fda call sites get it without changes. Idempotent operations (GET/HEAD/PUT/DELETE plus desired-state POSTs) retry on any transport failure and on HTTP 408/429/502/503/504. Non-idempotent operations (ingest, clock-advance, transactions) retry only when the first attempt provably never reached its target: a local connect failure, or a 503 PipelineInteractionUnreachable whose body carries the api-server's connect-phase wording. A Retry-After header overrides the computed backoff. This mirrors the Python SDK's RetryConfig, but gates statuses and timeouts on idempotency, which the Python SDK does not.
3. Pick the 502 retry wait by probing cluster health. Ports the Python SDK's cluster-aware 502 handling: before retrying a 502 the client probes /v0/cluster_healthz (200 only when every service is healthy). Healthy means the 502 was spurious and the retry runs immediately; unhealthy (e.g. an upgrade in progress) waits a flat unhealthy_backoff (90s per retry, so 270s total at the default 3 retries) instead of ramping exponentially.

Describe Manual Test Plan

• Ran crates/fda/test.bash end to end against the enterprise dev instance including the pclock clock-advance section that flakes in CI.
• Live retry smoke test: FELDERA_RETRIES=1 fda pipelines against an unreachable host logs the retry and then surfaces the error.
• Each retry behavior test was validated to fail with the behavior it guards reverted (hook impl removed, transport-retry condition narrowed, Retry-After ignored, health probe removed).

Checklist

• Unit tests added/updated
• Integration tests added/updated
• Documentation updated
• Changelog updated

Breaking Changes?

Mark if you think the answer is yes for any of these components:

• OpenAPI / REST HTTP API / feldera-types / manager (What is a breaking change?)
• Feldera SQL (Syntax, Semantics)
• feldera-sqllib (incl. dependencies fxp, etc.) (What is a breaking change?)
• Python SDK (What is a breaking change?)
• fda (CLI arguments)
• Adapters (including configuration)
• Storage Format / Checkpoints
• Others (specify)

Describe Incompatible Changes

• fda bench loses the --upload, --project, --branch, --start-point*, --benchmark-host, and --benchmark-token flags (BENCHER_* env vars).
• feldera-rest-api crate API: Client::new/Client::new_with_client take a new RetryPolicy argument, and the progenitor 0.11 Error enum replaces PreHookError/PostHookError with Custom. Retrying changes client behavior for consumers that relied on single-attempt semantics; construct the client with RetryPolicy::none() to opt out.