v5.0.0 Major Refactor: Move from Vue 3 + JSDoc to Svelte 5 + Typescript

[Sudo-Ivan]

Why? Better maintainability and opportunity to break down all those "god" files into components and more maintainable frontend. I am more experienced with Svelte and Typescript.

This will also feature large backend refactors to break down meshchat.py.

Using various open-weight models.

[kilo-code-bot]

[WARNING]: Budget prefix mdi- matches no emitted chunk — the CI budget check hard-fails

check-frontend-chunk-budgets.mjs fails the run ("No chunk matching prefix ...", exit 1) when no asset starts with the prefix. No build rule produces an mdi--prefixed chunk: the generated subset js/generated/mdiIconData.ts is statically imported by js/mdiIconNames.ts (line 4), so it merges into an app-/shared-async- chunk, and the only dynamic import (import("@mdi/js")) is captured by the vendor-mdi group in vite.config.mjs (which the longer vendor-mdi- prefix owns). Only vendor-mdi-* assets are produced, so pnpm run check:bundle-budgets in .github/workflows/frontend-build.yml fails on this entry. Remove it (the generated subset is already covered by the app- and shared-async- budgets).

---

Reply with @kilocode-bot fix it to have Kilo Code address this issue.

[kilo-code-bot]

[WARNING]: announce_count is never populated — the "most announced" sort silently degenerates

The sidebar's nodes come only from GET /api/v1/announces (via fetchNomadNodes) and WS announce merges (via mergeNomadAnnounceIntoNodes), and neither source carries announce_count: batch_convert_announces_to_api_dicts in announce_manager.py/lifecycle/announce_convert.py rebuilds each row from an explicit key whitelist that omits announce_count (likewise convert_db_announce_to_dict and the WS announce payload). So count is always 0 and most_announced sorts identically to newest_discovered. Populate the field in the API dict (or via a COUNT(*) aggregate) or drop the sort mode.

---

Reply with @kilocode-bot fix it to have Kilo Code address this issue.

[kilo-code-bot]

[WARNING]: blockedHashSet reads untracked GlobalState — blocked/unblocked state goes permanently stale

GlobalState is a plain Proxy with its own subscription list (js/GlobalState.ts), not a Svelte $state signal, so reading GlobalState.blockedDestinations inside $derived registers no dependency: this set is evaluated once and never recomputed. When a node is later blocked or unblocked (block-status-changed refetches and reassigns GlobalState.blockedDestinations), isBlocked(), the row highlighting, the context-menu block/lift label and bulkBlockNodes filtering all stay frozen until the component remounts. Subscribe to GlobalState and copy blockedDestinations into a $state (the established pattern, e.g. mirroring ConversationViewer.svelte's subscription approach) instead of reading the proxy in $derived.

---

Reply with @kilocode-bot fix it to have Kilo Code address this issue.

[kilo-code-bot]

[WARNING]: Document-level click listener fires after the toggle's focus, clearing the mobile search results

When the magnify/close toggle in MapToolbarControls.svelte is tapped, toggleMobileSearch() sets isMobileSearchOpen, then await tick() yields — a microtask checkpoint runs before the event bubbles on to document, so mobileSearchBar?.focus() fires onfocus (isSearchFocused = true, line 1680) before handleSearchClickOutside runs. The toggle button sits outside searchContainer/inlineSearchContainer, so insideSearch is false and the handler immediately resets isSearchFocused to false: showResults={isSearchFocused} never becomes true, the input is already focused (so re-tapping never re-fires onfocus), and typing only sets searchResults — the results list never renders until the user blurs and refocuses. Exclude the toggle button (or stop the click on it / check event.target against the toggle) in handleSearchClickOutside.

---

Reply with @kilocode-bot fix it to have Kilo Code address this issue.

[kilo-code-bot]

[SUGGESTION]: Toast messages bypass i18n with hardcoded English strings

The menu labels above use t("nomadnet.copy_address") / t("nomadnet.copy_nomad_link"), but the success toasts (line 63 and line 74: "Address copied to clipboard", "Link copied to clipboard") are raw English literals. Pass them through t() like the rest of the component.

---

Reply with @kilocode-bot fix it to have Kilo Code address this issue.

[kilo-code-bot]

[WARNING]: Check-then-fill race can pin a stale value in the config cache

get() releases _cache_lock for the fetchone() (line 28) and only re-acquires it to fill the cache (line 30). If thread A misses and fetches the row, thread B then runs set() (which writes the DB and pops the cache key), and a stale result can't be re-checked — thread A finishes set() first, then thread A's fill stores the pre-write value, which then stays cached indefinitely (no TTL) until the next write to that key. Re-check the cache inside the second lock before filling (or hold the lock across the fetch) so a concurrent set()/delete() invalidation can't be overwritten by an older read.

---

Reply with @kilocode-bot fix it to have Kilo Code address this issue.

[kilo-code-bot]

[SUGGESTION]: Promise.all couples peer loading to the contacts endpoint

Previously the two requests ran sequentially, so a /api/v1/telephone/contacts failure still left peers populated. Now one rejected request rejects the whole await, and the catch logs without setting either — a contacts-endpoint hiccup loses the palette's peer list too. Use Promise.allSettled (or keep the contacts fetch in its own catch) so each source fails independently.

---

Reply with @kilocode-bot fix it to have Kilo Code address this issue.

[kilo-code-bot]

CRITICAL: Decorated module-level symbols are corrupted by shared-module extraction.

_slice(lines, node.lineno, node.end_lineno) starts at the def/class line, but node.lineno for a decorated symbol points at the def, not the first @decorator. Two failures result:

1. The decorator (@lru_cache, @dataclass, @functools.cache, ...) is dropped from the emitted {stem}_shared.py, silently changing behavior of the extracted symbol.
2. The removal range (node.lineno, node.end_lineno) leaves the decorator lines behind in the rewritten file, where they now decorate whatever statement follows - usually a syntax error, occasionally a silent mis-attribution.

cmd_verify never compiles the shared module and never byte-checks shared symbols (only moved methods), so a split that mangles a decorated helper passes verify - violating the tool's stated lossless contract.

Fix: compute the decorated start the same way cmd_verify does for methods, e.g. _attached_comment_start(lines, min([node.lineno] + [d.lineno for d in getattr(node, 'decorator_list', [])])) for both the shared-body slice and the removal range.

---

Reply with @kilocode-bot fix it to have Kilo Code address this issue.

[kilo-code-bot]

SUGGESTION: max() over an empty sequence crashes for a module with no imports.

If the source module has zero top-level imports, the generator yields nothing and max() raises ValueError: max() arg is an empty sequence instead of a clear tool error.

Suggested change

	last_import_line = max(
	last_import_line = max(
	(n.end_lineno or n.lineno)
	for n in info["tree"].body
	if isinstance(n, (ast.Import, ast.ImportFrom))
	, default=0)

(With default=0 the insertion point becomes line 1, i.e. the top of the file, which is the correct spot when there are no imports.)

---

Reply with @kilocode-bot fix it to have Kilo Code address this issue.