Add subagent tool progressive disclosure

[Aaronontheweb]

Summary

• Seed child model requests from policy-visible Core tools.
• Bind search and load to each child private registry.
• Keep loaded deferred tools local to one child run.
• Exclude recursive delegation from child catalog and dispatch paths.
• Add count-only exposure diagnostics and deterministic lifecycle tests.

Stack

This PR depends on #2021. Merge the stack in order.

Deterministic evaluation evidence

• A child starts from the same policy-visible Core set as its parent, excluding recursive delegation.
• Against the fixed 200-tool synthetic catalog, the child initial footprint is lower than the frozen 202-definition, 133,816-byte baseline.
• Exact search, fuzzy suggestion, load, and direct dispatch cannot reveal or invoke recursive delegation or hidden tools.
• Loading one Deferred tool affects only that child run; a fresh child and a post-failure child do not inherit it.
• The exposure diagnostic contains counts only and omits names, paths, arguments, prompts, and schema bodies.
• No hosted behavioral run was used as evidence for this PR.

Validation

• Netclaw.Actors.Tests: 3,466 passed; one expected Windows-only skip.
• Focused post-rebase tests: 26 passed.
• Strict OpenSpec validation passed.
• Copyright header validation passed.
• Changed-file formatting passed.
• Slopwatch reported zero issues.
• Diff and BOM checks passed.