Follow up parallel lanes with measured capacity and deferred execution controls #1979

The concurrent lane runtime in agent-harness#1978 defaults to four lanes and clamps using host available memory and raw remaining ancestor-cgroup memory, with a configurable provisional 2 GiB per-lane budget. The budget is policy, not a measured child footprint; the 4 GiB seat limit is not used as a usage estimate.

Track the next adaptive-capacity increment:

• Measure child memory during the first real concurrent phase and recalibrate the provisional budget.
• Evaluate conservative reclaimable-cache credit, excluding dirty/writeback memory and avoiding double counting. Preserve raw readings and test the distinction between reclaimable cache and process usage.
• Incorporate usable CPU, load, memory pressure, swap and between-child admission without deriving worker count from core count alone.
• Retain default four, configurable ceilings, serial fallback with recorded reasons, and serialized heavy verification.

Other deferred parallel-lane items: direct-session execution gates, cross-phase defaults and pool cap, explicit stale-lane recovery, stronger process-tree quiescence evidence, Claude machine-verified ownership, and dynamic ownership claims after measurement. These remain separate from the skill-wording follow-up.