[FEATURE] Add prompt-based skill routing evaluation

[luongnv89]

Problem Statement

Users want to know which skill is actually selected when they give a certain kind of prompt.

Today, skill discovery is mostly static:

• browse/search skills
• inspect descriptions
• guess which one might be selected

But there is no practical way to evaluate real skill-routing behavior from a prompt and measure how stable or ambiguous that routing is.

This matters because LLM-based selection is stochastic. A single run does not tell us enough. We need repeated runs and measurable routing output.

Proposed Solution

Add a prompt-based skill routing evaluation feature.

Example workflow:

1. User provides a prompt
2. ASM runs that prompt through a selected tool/runtime multiple times
3. ASM records which skills are selected across runs
4. ASM reports the results as a routing/evaluation summary

Example command shape:

• asm eval-route --prompt "..." --tool openclaw --runs 20

or similar.

Suggested Output

For each evaluated prompt, ASM should report:

• selected skill(s)
• frequency across runs
• top-1 / top-k selection rate
• variance / ambiguity indicators
• tool/runtime used
• run count and evaluation settings

Optional future metrics:

• entropy / routing confidence proxy
• confusion between overlapping skills
• comparison across multiple tools/models

Why This Matters

This would turn skill selection into something measurable instead of guesswork.

It helps:

• users understand what skill will likely be used
• authors evaluate whether a skill is being selected as intended
• maintainers detect overlap and routing ambiguity
• ASM build real-world routing data for search, ranking, and trigger improvements

Alternatives Considered

• manually testing prompts one by one
• guessing from skill descriptions alone
• relying only on semantic search or trigger text matching

These approaches do not show real routing behavior under repeated usage.

Use Cases

1. A user wants to know which skill is most likely to be used for a task
2. A skill author wants to verify routing behavior for representative prompts
3. A maintainer wants to detect competing/ambiguous skills
4. ASM wants to build a benchmark-style dataset for skill routing quality

Additional Context

This is similar in spirit to recent tool-use / routing evaluation work:

• repeated-run evaluation matters because LLM routing is stochastic
• benchmark-style tool invocation evaluation is now common in agent/tooling research

So this should be designed as a lightweight routing-eval feature, not just a one-off debug command.

[luongnv89]

Deferred — see docs/SKILLGRADE_INTEGRATION_PLAN.md §1. Will revisit if demand signal appears.