Refactor describe_trust_system() into a more general load_skill() tool

[simonpcouch]

Related to #404.

Right now, describe_trust_system() is essentially an agent skill, but it felt silly to implement a skills system just for it. After #404, we will have another agent-skill-ish piece of prompting.

This is deliberately not a full agent skills implementation—no support for user-contributed skills.

metadata.topic becomes "Reading up on ..." in the user-facing title.

[github-actions]

Preview deployed to Connect (dogfood.team.pct.posit.it): https://dogfood.team.pct.posit.it/connect/#/apps/d7a36cae-8f27-448b-a478-61b81fbe3942/draft/378021

Deployed from commit 84be816.

[github-actions]

Preview root: https://posit-dev.github.io/commons/pr-405/

Python site preview: https://posit-dev.github.io/commons/pr-405/py/

Built from the latest commit on this branch. The R links in it point at the published R site, which no pull request rebuilds.

[github-actions]

Preview deployed to Connect (connect.staging.pct.posit.it): https://connect.staging.pct.posit.it/connect/#/apps/ad662e1b-5048-4acc-9ad7-f9478c92274e/draft/3413

Deployed from commit 84be816.

[simonpcouch]

Review from Claude Opus 5.5

Review (agent-written)

Reviewed against origin/main. A clean, well-scoped refactor: one skill format, one reader per language, and the model-facing description is pinned in tests/shared/tool-registration.json. Nothing blocking.

• minor (simplification) — yaml moves from Suggests to Imports, so the existing guards are now dead and can go:

  # R/data-dictionary.R
  rlang::check_installed("yaml")
  # R/catalog-snowflake.R
  rlang::check_installed("yaml", call = call)

• minor (tests) — test-skills.R and test_load_skill_returns_instructions_without_frontmatter assert the same behavior separately in each language, and nzchar() / assert body don't say much. What the tool returns (body with frontmatter removed, the Read up on answer trust title) is user-observable and implemented twice, so it could be a load_skill case in tests/shared/ instead. The unknown-name check can stay per language.

• minor (simplification) — read_skill() is a third copy of the frontmatter split in R, next to strip_frontmatter() in R/context-layer.R and read_review_document() in R/trajectory-review-log.R. The Python side already has strip_frontmatter() in _context_layer.py. Reusing those for the body would leave only the YAML parse new. Judgment call at this size.

• minor (clarity) — The PR description says the title becomes "Reading up on ...", but the result title is Read up on answer trust and the in-progress title is Reading up. The code follows the existing Searching/Searched convention, so it's just the description that's off.

No correctness issues found.

[simonpcouch]

skills.mov

[github-actions]

Cleaned up 3 preview bundle(s) on https://dogfood.team.pct.posit.it: 378013, 378016, 378021

[github-actions]

Cleaned up 3 preview bundle(s) on https://connect.staging.pct.posit.it: 3406, 3409, 3413