fix: call OnDiscard exactly once per item on encode failure

[acoshift]

Summary

• Finding 1 — double-discard on retry: flush previously called invokeOnDiscard during JSON encoding, but left the item in the buffer slice. On every HTTP retry the same item was re-encoded (failing again) and the callback fired a second time. Fixed by collecting encode failures in a local slice and only calling invokeOnDiscard after a successful HTTP response.
• Finding 2 — empty-body infinite retry: When every item in a batch was unencodable the encoded buffer was empty. flush still POSTed a zero-byte body; if the server rejected it retryFlush looped indefinitely, re-discarding items each iteration. Fixed by returning true early (and discarding all failures once) when buf.Len() == 0, skipping the HTTP request entirely.

Test plan

• TestIngest_EncodeError_DiscardedExactlyOnce_OnHTTPFailure — server fails twice then succeeds; verifies OnDiscard fires exactly once for the unencodable item despite multiple retries
• TestIngest_AllEncodeErrors_NoHTTPRequest — all items unencodable; verifies zero HTTP requests are made and OnDiscard fires once per item
• Existing TestIngest_JSONEncodeError_CallsOnDiscard still passes (happy-path encode error with successful HTTP)
• go test ./... passes

🤖 Generated with Claude Code