fix(registry): dedupe getAllCommands so aliased commands are not exported twice

[shoemoney]

config export-schema currently emits 19 tool schemas with only 17 distinct names:

$ mmx config export-schema | jq -r '.[].name' | sort | uniq -d
mmx_search_query
mmx_speech_synthesize

Duplicate function names are rejected by both the Anthropic and OpenAI tool APIs, so the exported schema cannot be handed to a model as-is without post-processing.

Cause. Two commands are registered under a second alias path:

'speech synthesize': speechSynthesize,
'speech generate':   speechSynthesize,   // same object
'search query':      searchQuery,
'search web':        searchQuery,        // same object

CommandRegistry.getAllCommands() walks the path trie and pushes node.command at every node that has one — so an aliased command is pushed once per path, not once per command. export-schema maps that list straight to generateToolSchema, and the duplicates land in the output.

Fix. Dedupe by object identity during the traversal. Callers of getAllCommands want the set of distinct commands, not the set of invocation paths.

Verified locally (bun run src/main.ts):

before   total entries: 19   unique names: 17   duplicates: {mmx_speech_synthesize: 2, mmx_search_query: 2}
after    total entries: 17   unique names: 17   duplicates: none

Alias routing is untouched — mmx speech generate --help and mmx search web --help both still resolve and exit 0. Only enumeration changes.

bun test: 385 pass / 8 fail, identical to the baseline on a clean checkout with this change stashed. Those 8 appear to be network-dependent (one logs Detecting region... failed) and are unrelated.

Happy to add a regression test asserting getAllCommands().length === new Set(getAllCommands()).size if you'd like one — I left it out because the repo's registry tests looked deliberately light, but say the word.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[NianJiuZst]

Thanks for the contribution! I verified the current head (8f352b6) against main: config export-schema changes from 19 entries with 17 unique names to 17/17, while the speech generate and search web aliases continue to resolve correctly.

The identity-based deduplication is narrowly scoped and matches the current use of getAllCommands(). The exact-head CI is green, and local validation passed all 449 tests, typecheck, and the development build; lint has no errors (only one unrelated existing warning).

I found no blocking issues, and this looks ready to merge. A direct regression assertion that exported schema names remain unique would be a nice follow-up, but I do not consider it blocking.

Thanks again for the clear explanation and focused fix!