Translate only the literal path of a ~q route

[rubas]

Fixes two bugs in how ~q translates a route, with one change.

Before

Localize.put_locale(:fr)
~q"/users?next=/users"
#=> "/users?next=/users_fr"
~q"/users/#{String.upcase("users")}"
#=> "/users_fr/USERS_FR"

Problem

• A / inside the query hides the real end of the path. The path stays untranslated, and a query value is translated instead. Literal text after an interpolated query value (?page=#{page}&next=/users) is also translated.
• String literals inside #{...} code are translated too. A literal that matches a routes msgid (a default value, a map key, a branch result) is silently replaced at compile time.

After

Localize.put_locale(:fr)
~q"/users?next=/users"
#=> "/users_fr?next=/users"
~q"/users?page=#{2}&next=/users"
#=> "/users_fr?page=2&next=/users"
~q"/users/#{String.upcase("users")}"
#=> "/users_fr/USERS"

Details

• Cause: Localize.Routes.translate_path_now/4 used Macro.prewalk/2, which also visits the code inside each interpolation. translate_segment_parts/4 split each literal on / first and only looked for ? or # in the last piece.
• Change: translate_path_now/4 walks only the top-level <<>> segments with Enum.map_reduce/3. Each literal is split once at the first ? or #; only the part before it is translated, and every later segment is left as is. Interpolation segments are never walked. ~q, path_for/2 and url_for/2 always pass a <<>> node (normalize_route_ast/1). The now unused catch-all clause of merge_literal_segments/1 is removed (the compiler flagged it).
• Tests in test/sigil_q_test.exs, both fail on main: sigil_q translates the path but not a path in the query and sigil_q does not translate string literals in interpolated code.
• Checks: mix test (269 passed), mix compile --warnings-as-errors, mix format --check-formatted, mix credo --strict (only the existing Sigil_q.InterpolationTest name note).

---

AI disclosure: AI models helped us find this issue, write the change, and review it.