feat(core): add gettext_noop marker and translation-catalog fixture

[hamza-56]

Part of: Sparkth UI, emails, API errors, and AI-generated content are English-only

First layer of the i18n stack (#601 → #602 → #603 → #604 → #606 → #607) that follows up the translation foundations from #590.

What

Adds the missing marking primitive and the shared test tooling that the string-marking PRs stacked on top rely on: a deferred gettext_noop marker for strings that must stay plain str, and a fixture for asserting on real translations.

Changes

• feat(core): gettext_noop marking function, exported through sparkth.lib.i18n and extracted via -k gettext_noop; for literals stored in plain-str fields (the DisplayInfo/SidebarEntry dataclasses embedded in Pydantic response models, where a LazyString cannot go), translated later by passing the stored value through gettext() at the rendering boundary
• test(core): translation_catalog fixture in sparkth.lib.testing — compiles single-message catalogs onto the LOCALE_DIRS hook so any test (core or plugin) can assert on an actual translation
• test(core): shipped_locale_dirs session fixture detaching the shipped catalog directories for the whole suite, so locally compiled .mo files (make i18n.compile) never leak into assertions written against the English source fallback
• docs: translations guide gains the fourth marking case and the new testing workflow

How to Test

1. uv run pytest tests/core/i18n (24 passed)
2. make i18n.compile && uv run pytest — the suite stays green with compiled catalogs present (the detachment fixture at work); before this PR the compiled-catalog state broke test_gettext_returns_the_source_when_the_locale_has_no_catalog
3. make mypy and make lint.backend are clean

Notes

No behavior change outside tests: gettext_noop has no call sites yet (they arrive in #602/#603).

This PR description was written with the assistance of an LLM (Claude).

[hamza-56]

@claude-review

[claude]

Claude finished @hamza-56's task in 2m 28s —— View job

---

Claude Code Review

• Read the PR diff and changed files
• Review correctness, tests, and docs
• Post review feedback

Reviewed all 7 changed files against origin/main. This is a clean, well-scoped foundation PR — the marking primitive and the test tooling both land with docs and tests, no behavior change outside tests (as advertised). Nice work.

Overall assessment

Approve-worthy. No blocking or correctness issues found. A couple of low-priority nits below.

What's good

• gettext_noop is a minimal, correct identity marker, exported symmetrically through core.i18n.__init__ → lib.i18n and covered by tests at both the core and façade layers.
• shipped_locale_dirs is a genuinely nice fix: detaching the shipped catalog dirs for the whole session means a developer's locally-compiled .mo files can't leak into English-fallback assertions. This closes the real footgun the PR body calls out (test_gettext_returns_the_source_when_the_locale_has_no_catalog breaking under compiled catalogs).
• translation_catalog correctly sidesteps the _load_catalogs lru_cache(maxsize=None) staleness trap by compiling each message into a fresh tmp_path subdir — every registration mutates the (locale, dirs) cache key, so no invalidation is needed. The Protocol return type and per-registration teardown are tidy.
• Makefile: -k gettext_noop added to i18n.extract, which both i18n.init and i18n.update depend on, so extraction stays consistent across all targets. -k augments (not replaces) pybabel's default keyword set, so _/gettext/lazy_gettext still extract. 👍

Suggestions (Low)

1. docs/guides/translations.md:127 — the testing example is a plain def but uses await client.post(...). Under asyncio_mode = "auto" this needs async def; as written it's a SyntaxError if copied. (inline comment + suggestion)

2. tests/core/i18n/test_translate.py:83 — test_gettext_noop_returns_the_message_unchanged requests catalog_dir, but gettext_noop never consults a catalog, so the fixture is dead weight and slightly muddies intent. (inline comment)

3. No consumer for translation_catalog yet (non-blocking) — the new fixture has no test exercising it in this PR (the two gettext_noop tests reuse the existing catalog_dir). That's reasonable for stacked tooling that feat: mark core and api user-facing strings for translation #602/feat(plugins): mark plugin user-facing strings for translation #603 will use, but it means CI doesn't validate the fixture's happy path here. Consider having one of the new gettext_noop tests route through translation_catalog to both cover it and retire the duplicated catalog_dir/write_catalog logic.

Notes

• No frontend files touched — React anti-pattern checks N/A.
• No security, async-safety, or exception-handling concerns; no bare except.
• TDD respected: tests accompany the new marker and the façade export.
  · i18n-gettext-noop

[claude]

The example body uses await client.post(...) but the function is declared with a plain def. The suite runs asyncio_mode = "auto", so this test would need to be an async def to actually await. As written, copying this pattern produces a SyntaxError (await outside an async function).

Suggested change

	def test_detail_is_translated(client, translation_catalog):
	async def test_detail_is_translated(client, translation_catalog):

[claude]

Minor: this test requests catalog_dir, but gettext_noop never consults a catalog (it returns its argument verbatim), so the fixture is unused here — it registers a Spanish catalog that has no bearing on the assertion. Dropping the parameter makes the test's intent (pure identity, independent of any catalog) clearer:

def test_gettext_noop_returns_the_message_unchanged() -> None:
    with locale_context("es"):
        assert gettext_noop(TRANSLATED[0]) == TRANSLATED[0]