Remove the Gemma hard memory ceiling

[8Dionysus]

Summary

• clear inherited hard memory caps for the selected Gemma 4 E2B llama.cpp lane with an explicit unlimited override
• keep the 4 GiB soft reclaim reservation and owner-native 600-second idle sleep
• align the existing runtime policy, tuning docs, model card, contract test, changelog, and repo-local KAG indexes

Why

The Gemma overlay previously replaced an inherited 12 GiB candidate ceiling with an 8 GiB ceiling. This important inference lane already has a measured owner-native release path: idle sleep unloads model and KV residency, while real requests wake it. A private hard ceiling can therefore force reclaim without providing the useful lifecycle action.

This PR changes source truth only. It does not sync the deploy mirror or recreate the running container.

Validation

• python scripts/validate_stack.py
• focused compose and service-policy tests
• full intel-federation source render: mem_limit=0, mem_reservation=4g, native sleep command retained
• pinned repo-local KAG index-family parity check
• all validation launched through abyss-machine resource launch; largest peak was 221.3 MiB with zero swap

Known base debt: python scripts/ci_gate.py --mode source-fast stops on pre-existing required-snippet drift in mcp/services/aoa-kag-mcp/AGENTS.md; this PR does not touch that owner surface.