LATENESS violation in as-of-like inequality join + arg_max #6829

Describe the bug
Hi all, I think I've found a case which violates the LATENESS contract. An as-of lookup written as an inequality join plus ARG_MAX silently drops an output row when LATENESS is declared on the timestamp columns.

This is not causing me pain; I can use the ASOF JOIN for my usecase. But I found it during some testing and thought I'd flag it. You guys rock! 🪨 ❤️

This PR contains a minimal repro where the LATENESS contract is not violated by inputs but outputs are different than an identical SQL with LATENESS omitted.

Related issues/links from my agent:

• feldera/feldera#6690 [SQL] Output of input_map_with_waterline should not be GC-ed using retain_values #6690 / feldera/feldera#6693 [SQL] GC of tables with keys and LATENESS should never use retain_values, but can use retain_keys when keys have LATENESS #6693(fix merged 2026-07-27, already in the pinned commit): upstream established that integrate_trace_retain_values for LATENESS GC is unsound because it "could delete keys that should be permanently frozen," and switched the input_map_with_waterline / PRIMARY KEY path to integrate_trace_retain_keys. This report is the same root cause in an uncovered codepath (the history join input of an inequality-join + ARG_MAX plan).
• feldera/feldera#4631 [SQL] GC output of input_map_with_waterline #4631 (merged 2025-08-26): introduced the retain_values waterline GC later flagged as unsound.