perf: bound tool I/O and trim tool surface

[huytg2610]

Summary

• Bound Read, WebFetch, and ReadPDF input work to reduce memory and latency spikes.
• Add compact tool profiles; the standard profile sends 14 of 38 schemas (61% smaller schema payload).
• Keep legacy saved configurations on full, add cache authorization preflight, config-aware cache keys, and concurrent-safe LRU access.

Validation

• 157 focused regression tests passed.
• 16 daemon integration tests passed.
• Three independent review passes completed before push.