fix(extensions): list commands the way the active agent registers them

[tkv00]

Description

specify extension add ("Provided commands") and specify extension info ("Commands:") list an extension's commands. Both special-cased Cline and Forge by agent key, and every other agent got the manifest's dotted name. For integrations that register hyphenated names, the listings show commands the agent does not have:

Project	Listed on main	Registered
--integration codex	speckit.demo.hello	.agents/skills/speckit-demo-hello/SKILL.md
--integration claude	speckit.demo.hello	.claude/skills/speckit-demo-hello/SKILL.md
--integration junie	speckit.demo.hello	.junie/commands/speckit-demo-hello.md
--integration copilot (skills, the default)	speckit.demo.hello	.github/skills/speckit-demo-hello/SKILL.md

The per-agent branches grew one agent at a time (#2508 for Cline, #3669 and #3744 for Forge). Of the 41 registered integrations (other than generic), 25 register hyphenated names, and the listings handled 2.

This change replaces both branches with one helper in extensions/_commands.py. The helper asks the active integration for its command separator in the project's mode: invoke_separator_for_mode() with the persisted ai_skills flag, the hook whose docstring names extension registration paths. It then hyphenates with ExtensionManager._skill_name_for_command, which skill registration already uses. This follows #4910, which derived the init next-steps names from the integration for the same reason.

What stays the same:

• Cline and Forge print the same names. Manifest command names must match ^speckit\.([a-z0-9-]+)\.([a-z0-9-]+)$, and for such names their formatters and the shared helper return the same string.
• Agents that register dotted names, such as Gemini and Copilot with --commands, keep the dotted name.
• An unknown or unreadable ai value falls back to the manifest name, as before.

The helper relies on a - command-mode separator meaning hyphenated registration. A new registry test checks this for every registered integration other than generic (all 41 hold today).

Related: #4954 reports a similar mismatch in hook "To execute" lines for Codex. That code path is separate and is not changed here.

Testing

• New test_extension_install_listing_matches_registered_command_names and test_extension_info_listing_matches_registered_command_names share LISTED_COMMAND_NAME_CASES (8 agent and mode cases). On main, 8 fail (claude, codex, junie and copilot-skills, for each command) and 8 pass. With the fix, all 16 pass.
• New test_command_mode_separator_matches_registered_names in tests/integrations/test_registry.py (41 cases). It also passes on main; it guards the assumption the helper relies on.
• tests/specify_cli/extensions, tests/integrations, tests/test_extensions.py and tests/test_installed_list_json.py show no new failures compared with main.
• Full suite: I compared the failing tests with main in the same local environment. The branch adds no new failures. One timing-dependent workflow test failed once and passed 5 of 5 reruns.
• CLI: I created projects with specify init --integration <codex|claude|junie|copilot|gemini> and with copilot --integration-options="--commands". For a sample extension, extension add --dev and extension info now list speckit-demo-hello for codex, claude, junie and copilot, and still list speckit.demo.hello for gemini and copilot --commands.
• uvx ruff@0.15.0 check src tests passes.

Tests ran with the project virtualenv's python -m pytest after uv sync --extra test.

• Tested locally with uv run specify --help
• Ran existing tests with uv sync && uv run pytest
• Tested with a sample project (if applicable)

AI Disclosure

• I did not use AI assistance for this contribution
• I did use AI assistance (fill in the disclosure below)

AI disclosure: Prepared with Claude Code (model: Claude Opus 5.5) on behalf of @tkv00, running autonomously under step-by-step direction. The contributor chose the topic and the fix approach and approved each step. Claude Code found the issue, wrote the code, tests, documentation line, commit messages and this description, and ran the tests. AI generated the code, tests and text. No human reviewed the diff line by line before submission.