Allow explicitly larger compact input exports without changing default bounds

[CaelmBleidd]

The 30-second calibration generated 13–14k inputs but retained only 3956 under the 2 MiB export limit, making complete class-coverage comparison ineligible. Allow an explicit input-export byte limit up to 64 MiB while keeping its default at 2 MiB and the invocation bound at 100k. Increase the private stdout allowance only by the extra bytes explicitly requested for that export; ordinary requests keep the existing 4 MiB stdout, 4 MiB input and 64 KiB stderr limits. The shared transport API and timeout behavior are unchanged.

The experimental driver accepts the same explicit bound. Sampling quotas remain independent of capture limits, and omissions still disqualify complete coverage. No spooling, sessions or persistent capture service is added. A small style correction from the current coordinator branch is included in the existing guided test.

Stacked directly on #597 b09a010f8, independently of the numeric coordinator.

Validation: a real native/backend regression returns over 4 MiB of compact inputs alongside coverage and point observations, preserves 600 inputs including lone UTF-16 surrogates with an explicit 8 MiB limit, and demonstrates retained default overflow on the same values. Eight input-export tests, the generation-deadline regression, process-client boundary tests, 220 Node tests, nine Python tests and main/test Detekt passed. Repeat actual-budget calibration with the new frozen limit before a broad comparison.