Reduce node-state serialization overhead

[the-code-learner]

Scope

First profile-driven Zephyr performance experiment, stacked on the Consensus & Performance Lab PR.

The canonical 7-validator profile shows the current persistence hot path dominates before P-256 verification:

• ledger.(*Store).writeState: ~58% cumulative CPU
• encoding/json.MarshalIndent: ~55% cumulative CPU
• MarshalIndent: ~41% of flat allocations in the profiled process
• Store.Accept: ~90% cumulative mutex delay

Experiment

Replace machine-only state.json serialization from json.MarshalIndent to compact json.Marshal while preserving:

• the same JSON semantics and schema
• the same state normalization
• the same chain binding
• the same atomic temp-file/fsync/rename persistence path
• the same owner-only file permissions
• the same restart/load behavior

Measurement

The PR temporarily runs baseline and optimized canonical 7-validator benchmarks in the same GitHub-hosted job with the same runner, workload and -benchtime=5x, then verifies go vet ./..., the full Go suite and the Consensus & Performance Lab before committing the experiment.

This is deliberately not the journal/checkpoint redesign yet. It isolates how much of the current bottleneck is purely pretty-printed full-state JSON serialization before we make a larger storage-architecture change.