perf(skills): speed up snapshot prompt builds

[dexhunter]

Summary

• Build the local skills snapshot manifest with one directory walk instead of separate index scans.
• Avoid importing gateway.session_context during normal CLI prompt startup.
• Reuse a direct platform-list matcher for snapshot entries and cover it with existing platform-gating tests.

Fixes #3356.

Benchmark

I ran autoresearch with Weco against a strict synthetic skills snapshot harness for the remaining first-turn skills/frontmatter bottleneck.

Weco dashboard: https://dashboard.weco.ai/share/DHk95734hgGIOpYoaXJxh-Vj6TDgWZA7

The harness builds 700 synthetic skills, seeds the disk snapshot, then measures fresh build_skills_system_prompt child-process calls. It also checks prompt hash stability, platform gating, tool fallback/requirement filtering, and snapshot invalidation after edit/add/delete.

Case	hermes_skills_snapshot_prompt_ms
Fresh baseline	69.07 ms
This patch	12.21 ms

That is about an 82% reduction on the measured snapshot prompt-build path.

Validation

• uvx ruff check --fix .
• uv run python -m py_compile agent/prompt_builder.py agent/skill_utils.py tests/agent/test_skill_utils.py
• uv run --with pytest --with pytest-timeout python -m pytest tests/agent/test_skill_utils.py::TestSkillMatchesPlatformTermux tests/agent/test_prompt_builder.py -q
  • 171 passed
• uv run python ~/osspr/weco_metric_scripts/NousResearch__hermes_agent_skills_prompt_snapshot/evaluate.py --repo-dir . --iterations 7
  • hermes_skills_snapshot_prompt_ms: 12.2136902064085

[kshitijk4poor]

Merged via #61131 — your commit was cherry-picked onto current main with your authorship preserved (now e21ba91 on main). Thanks for the contribution!

Validation that went into the salvage PR: manifest output byte-identical to main on a nested/support-dir/excluded-dir fixture, ~5.1x faster builds on a 300-skill fixture (24.9ms -> 4.9ms), and an AUTHOR_MAP entry so release attribution resolves to you. Closing this original in favor of the merged salvage.