Tool manifest bloat: prefer trigger-scoped tool exposure (split from #204)

[mostlydev]

Split out from #204 §I to keep the cursored-context PR focused.

Context

The Tiverton effective-context capture (2026-04-29T05:04:01Z, turn_count: 1) showed:

```text
tools count: 43
tools JSON total: 37853 chars
```

That is ~38 KB of tool schema in every request, even for a simple Discord context question that touches none of those tools. The system+tools+memory grand total was 81 KB per turn. Tool manifest is not the root cause of the first-system rewrite issue (#204), but it is a real lever for prompt size and provider latency.

Hypothesis

Tools should be trigger-scoped or role-scoped at compile time so only the relevant subset is exposed for a given invocation type. For example:

• A Discord-mention turn does not need market-data tools.
• A scheduled claw api schedule invocation might need the full set.
• A managed MCP sidecar's tools are only relevant when that sidecar is reachable.

Possible directions

1. Compile-time trigger filter in claw.describe — declare triggers: [discord, schedule] per tool, expose only intersecting tools per invocation.
2. Lazy tool manifest — initial request carries a compact manifest digest; tools are loaded on demand via a separate endpoint when the model calls one. Probably too much complexity for v1.
3. Per-agent tool allowlist in pod YAML — operator declares which tools each agent gets. Operator-driven, no model-driven dynamism. Cheapest option.

Open questions

1. Is the right knob agent-level, trigger-level, or per-skill?
2. Does nanobot support a compact tool digest pattern, or is the full schema always required?
3. How does this interact with managed MCP sidecars (Support MCP sidecar services as managed tools through cllama #177/Provide first-class stdio MCP wrapper for managed MCP sidecars #179) where tool sets are discovered at compile time?

Out of scope

• Lazy/on-demand tool loading (probably v2 if option 1 isn't enough).
• Cursored Discord context (cllama should inject Discord live context as cursored append-only turn deltas, not rewrite the first system prompt #204).
• Dispatch fallback (separate issue split from cllama should inject Discord live context as cursored append-only turn deltas, not rewrite the first system prompt #204 §H).

Predecessor

#204 ships first. The hash telemetry added in #204 makes the impact of any tool-manifest change measurable.