POK-87: Make npm agent uninstall registry match runtime behavior

[saheljalal]

Summary

• Fix resolve_uninstall_cmd() precedence: explicit uninstall_cmd (including None) now wins over npm version_check derivation.
• Remove misleading "uninstall_cmd": None from twelve npm agents; npm uninstall is derived from version_check when no explicit command is set.
• Add regression tests for both behaviors.

Closes POK-87

Test plan

• test_aikit_npm_agent_uninstall_derived_from_version_check
• test_aikit_explicit_uninstall_cmd_none_blocks_npm_derivation
• test_aikit_goose_uninstall_uses_brew_only_for_homebrew_install

[saheljalal]

Review — POK-87: npm agent uninstall registry

Verdict: Approve. Small, correct, and well-tested. The registry is now self-consistent and I confirmed runtime behavior is genuinely unchanged for all 31 agents.

What I verified

• Precedence flip is behavior-neutral. Scanned all 31 agents: exactly 12 are npm-version_check agents, and none retains an explicit uninstall_cmd. So no current agent hits the new "explicit wins over npm" branch with a value that differs from the old npm derivation — every agent resolves to the same command as before. The flip changes intent/semantics, not output.
• No stale key reads. uninstall_cmd is only read through resolve_uninstall_cmd() (vendor-cleanup scan + uninstall loop). Removing the 12 dead None rows can't KeyError — access is via in / .get().
• None path still handled downstream. Explicit None short-circuits to "no automated uninstall — remove manually", which the loop already prints; the new test_..._none_blocks_npm_derivation locks it in.
• Tests green. Full tests/test_tools_characterization.py passes (153), including both new regression tests.
• Version bump (1.15.5) and CHANGELOG entry are consistent and dated today.

Nit (optional, non-blocking)

• Consider one more test asserting an explicit non-None uninstall_cmd on an npm agent wins over the derived npm command — that's the actual precedence guarantee this fix introduces, and nothing in the current registry exercises it, so a future edit could regress it silently. Purely defensive; fine to skip.

Nice call choosing to both drop the dead None rows and honor an explicit uninstall_cmd — self-consistent registry without losing the escape hatch.