perf(subagents): bound the spawn tool schema

[tt-a1i]

Summary

• replace the unbounded subagent_spawn role prose with a deterministic selection projection: role name, bounded purpose, capability class, and default effort
• keep every effective role name in the strict dynamic enum, while moving full lifecycle and precedence guidance to the Subagents Skill
• use one canonical model-facing spawn surface for runtime registration and budget tests
• enforce UTF-8 byte budgets for each purpose, the role directory, the default roster, and a maximum legal 128-role roster

Closes #62.

Design research and peer comparison: #84.

Measured surface

Metric	Before	After	Change
default subagent_spawn surface	3,909 bytes	2,472 bytes	-36.8%
constructed maximum legal 128-role roster	241,423 bytes	14,029 bytes	-94.2%
resident prompt guidelines	875 bytes	457 bytes	-47.8%

All 128 role names remain valid enum values. The parent-facing role directory is capped at 4 KiB and reports omitted summaries without weakening runtime validation.

Representative model feature probe

One non-lockstep feature probe used seal/deepseek-v4-flash-0731-baidu with the same explicit read-only delegation prompt against base af950d5 and this branch.

• both versions selected explorer and successfully called subagent_spawn
• first-request input: 15,766 -> 15,345 (-421, -2.7%)
• spawn-request input: 16,217 -> 15,800 (-417, -2.6%)
• three recorded parent turns: 49,697 -> 48,604 total tokens (-1,093, -2.2%)

This is a feature probe, not a general quality benchmark. It establishes that the compressed projection preserved role selection in a representative model and reduced the request surface in the expected direction. An initial CLI allowlist attempt was invalid because the dynamic tool was not activated; it was aborted and excluded.

Validation

• subagent-focused tests: 37/37
• bun run check
• Node tests: 814/814
• Vitest: 30/30
• git diff --check

The existing 12 Effect type warnings in extensions/file-search remain unchanged.