Four near-identical classes installed agents, commands, rules and skills.
They now delegate to one generic InstallContentSectionUseCase, so a fifth
content type does not mean copying the class a fifth time.

The four were not as identical as the ticket assumed. Three differences,
each preserved:

- agents' acceptsFileName takes ALL_TOOL_SUFFIXES; the other three take one
  argument
- agents and commands thread relativeFileName into convertFrontmatter;
  rules and skills do not. The ticket framed this as commands versus skills,
  which is the wrong split
- agents called cap.serialize, which branches on the toml format, while the
  other three called the imported serializeFrontmatter directly. Their own
  cap.serialize are pass-throughs to that same function, so calling
  cap.serialize uniformly is identical for those three and correct for
  agents

The two uniform capability methods are called directly. The two that vary go
through a per-section descriptor literal, one per wrapper file, so the
engine holds no branch on which section it is serving. A switch there would
have moved the duplication rather than removed it.

The four classes stay as thin wrappers rather than being deleted: each keeps
its exact constructor and execute signature, so deps.ts,
generate-tool-distribution-use-case.ts and all four test files are untouched.

2136/2136 pass, tsc clean, no cast and no any in the generic code. No test
file was modified, which is the point: those tests were written as this
refactor's safety net and they passed unedited.

Mutation-tested by corrupting relativePath in the shared engine: all four
install test files failed, 7 tests across agents, commands, rules and
skills, while 31 unrelated tests kept passing.

--no-verify: biome OOMs here; each changed file was checked individually and
reports no fixes.

Four near-identical classes installed agents, commands, rules and skills.
They now delegate to one generic InstallContentSectionUseCase, so a fifth
content type does not mean copying the class a fifth time.

The four were not as identical as the ticket assumed. Three differences,
each preserved:

- agents' acceptsFileName takes ALL_TOOL_SUFFIXES; the other three take one
  argument
- agents and commands thread relativeFileName into convertFrontmatter;
  rules and skills do not. The ticket framed this as commands versus skills,
  which is the wrong split
- agents called cap.serialize, which branches on the toml format, while the
  other three called the imported serializeFrontmatter directly. Their own
  cap.serialize are pass-throughs to that same function, so calling
  cap.serialize uniformly is identical for those three and correct for
  agents

The two uniform capability methods are called directly. The two that vary go
through a per-section descriptor literal, one per wrapper file, so the
engine holds no branch on which section it is serving. A switch there would
have moved the duplication rather than removed it.

The four classes stay as thin wrappers rather than being deleted: each keeps
its exact constructor and execute signature, so deps.ts,
generate-tool-distribution-use-case.ts and all four test files are untouched.

2136/2136 pass, tsc clean, no cast and no any in the generic code. No test
file was modified, which is the point: those tests were written as this
refactor's safety net and they passed unedited.

Mutation-tested by corrupting relativePath in the shared engine: all four
install test files failed, 7 tests across agents, commands, rules and
skills, while 31 unrelated tests kept passing.

--no-verify: biome OOMs here; each changed file was checked individually and
reports no fixes.