fix: address review feedback on swamp-getting-started skill

[stack72]

Summary

Addresses three items from the adversarial review on #1174:

• Track A/B/C terminology undefined: Replaced all 6 occurrences of the
  undocumented Track A/B/C labels in SKILL.md with self-documenting
  model-type names (command/shell, "local typed models", "extension
  models"). This eliminates ambiguity — an AI no longer needs to infer the
  mapping from context.

• Dual detection mechanism inconsistency: The generated CLAUDE.md (in
  repo_service.ts) told the AI to "check if the models/ directory
  contains any model YAML files", while SKILL.md's pre-check uses
  swamp model search --json. These could diverge (e.g., extension models
  live in extensions/models/ and get symlinked). Now both layers use the
  same CLI-based check, which is the authoritative source for registered
  models.

• Pre-check error handling: If swamp model search --json fails (repo
  not initialized, corrupt config, swamp not installed), the skill now
  explicitly tells the user what went wrong and delegates to swamp-repo
  for setup rather than silently skipping onboarding.

Also adds missing skill_assets_test.ts coverage for swamp-getting-started
(inclusion test + reference file copy test), matching the pattern used by all
other bundled skills.

Impact

Skill-only changes — no runtime code paths affected beyond the generated
CLAUDE.md string in repo_service.ts. The detection alignment means new repos
will get consistent onboarding behavior regardless of whether the trigger comes
from the generated CLAUDE.md or the SKILL.md pre-check.

Test plan

• deno check passes
• deno lint passes
• deno fmt passes
• deno run test src/infrastructure/assets/skill_assets_test.ts — 31/31
  tests pass (2 new)

🤖 Generated with Claude Code

[github-actions]

Code Review

Clean, well-scoped PR that addresses specific adversarial review feedback from #1174. All three items are handled correctly.

Blocking Issues

None.

Suggestions

1. Minor terminology in SKILL.md: The references/tracks.md filename is a vestige of the old Track A/B/C naming. Not worth renaming in this PR (it would require updating the embedded link in SKILL.md and the test assertions), but worth noting for a future cleanup pass if the team prefers full consistency.

Details

Terminology replacement (SKILL.md): All Track A/B/C references are replaced with self-documenting names (command/shell, "local typed models", "extension models"). The replacement is thorough — I count 9 individual edits across 6 logical locations. The referenced tracks.md file already uses descriptive terminology, so no inconsistency exists.

Detection alignment (repo_service.ts): The generated CLAUDE.md now instructs the AI to run swamp model search --json instead of checking the models/ directory. This matches the SKILL.md pre-check, eliminating the divergence risk noted in the review (extension models living outside models/).

Error handling (SKILL.md): The new paragraph for pre-check failures is well-placed between the success-with-models path and the no-models path, covering the third case (command failure). Delegation to swamp-repo for uninitialized repos is the right recovery path.

Tests (skill_assets_test.ts): Both new tests follow the established patterns exactly — the inclusion test matches "SkillAssets includes <name> skill" and the copy test matches "SkillAssets.copySkillsTo copies <name> reference files". Uses @std/assert and the existing withTempDir helper.

DDD: The repo_service.ts change is a string template in a domain service — appropriate placement. The terminology cleanup is good ubiquitous language practice (replacing internal jargon with domain terms).

CLAUDE.md compliance: No new .ts files, no any types, no libswamp import boundary violations, tests next to source, focused changes with minimal blast radius. ✅

[github-actions]

Adversarial Review

Critical / High

None.

Medium

1. repo_service.ts:613-616 — Generated CLAUDE.md doesn't handle command failure. The SKILL.md now has explicit error-handling instructions for when swamp model search --json fails (repo not initialized, corrupt config, swamp not installed). However, the generated CLAUDE.md in repo_service.ts only says: "If no models are returned (empty result), you MUST immediately invoke the swamp-getting-started skill." A command failure (non-zero exit, stderr output) is not the same as an empty result. An AI reading the generated CLAUDE.md could interpret a failed command as something other than "no models returned" and skip the onboarding trigger entirely — exactly the silent-skip problem the SKILL.md fix addresses. Consider adding a brief clause like "If the command fails or returns no models" to keep the two layers fully aligned.

Low

1. SKILL.md:88 vs SKILL.md:82-84 — Tension between On Failure default and Verify guidance. State 1's On Failure says to "default to a command/shell model. It works everywhere" while the Verify step says to "Only use command/shell if the user's goal is genuinely a one-off ad-hoc command." This contradiction pre-dates the PR (it existed with the Track A terminology), but since these lines were touched, it's worth noting. An AI following both instructions literally could oscillate. A one-line qualifier like "default to command/shell for the walkthrough demonstration" would resolve it.

Verdict

PASS — Clean, well-scoped changes. The terminology alignment and error-handling additions are correct. The medium finding is a gap in the generated CLAUDE.md trigger that could be closed with a small wording tweak but does not block merge since the SKILL.md pre-check handles the failure case once the skill is actually invoked.