generate: report tools declared but not referenced by any agent

[LuisMend12]

Summary

• Generator.generate() only looked at tool_ids collected by walking the agent graph (_collect()), so a tool declared in the top-level tools: registry but never added to any agent's tools: list was invisible in agentctl generate output — no stub (correct, nothing uses it) but no mention either. A forgotten reference looked identical to an intentional one.
• Added SystemSpec.declared_tools, populated from the full top-level tools: registry during YAML parsing (previously this registry was only used transiently inside the parser to resolve per-agent tool references, then discarded).
• Generator.generate() now diffs declared_tools against the referenced tool_ids and reports the difference on a new GenerateResult.ignored field.
• agentctl generate surfaces it:

    ignore  add_new_user  (declared but not referenced by any agent)
  

Fixes #67

Test plan

• New test tests/parsers/test_declared_tools.py — confirms the parser populates SystemSpec.declared_tools (ids + descriptions) from real YAML, using the exact repro from the issue
• New tests in tests/generate/test_plugins_manifest.py — Generator.generate() reports an unreferenced declared tool in result.ignored (and not in created/no stub file), and reports nothing when all declared tools are referenced
• Ran tests/parsers, tests/generate, tests/core — all pass except 5 pre-existing failures unrelated to this change (a Windows-only path-separator assertion bug: result.created.count("plugins/plugins.toml") compares a forward-slash literal against str(Path.relative_to(...)), which yields backslashes on Windows — reproduces identically on main before this change)

[Asaf-prog]

Thanks for the contribution — the implementation direction is correct, and preserving the full top-level tool registry as declared_tools is a clear approach.

However, PR #74 is already open for the same issue and implements the same behavior, including the ignored result, CLI output, and regression coverage.

PR #74 also handles the ignored-only CLI case more accurately. In this PR, the command can currently print:

ignore add_new_user (declared but not referenced by any agent)
✓ Nothing to generate — all stubs already exist.

That summary is misleading because the ignored tool's stub does not exist by design. PR #74 uses a more accurate summary and includes an end-to-end CLI regression test.

To avoid maintaining and merging two competing implementations for #67, I think we should continue with #74 and close this PR as a duplicate.

The declared_tools naming used here is clearer, though, and is worth carrying over to #74.

Thank you again for taking the time to work on this.

[LuisMend12]

Totally agree, thanks for checking — didn't see #74 was already in flight. The ignored-only summary bug you caught there is a good catch, glad it's covered. Appreciate you calling out the declared_tools naming, hope it's useful if you revisit that field later.