fix(translate): retry a sentence batch whose reply cannot be read

[tnalbant]

What changed

The sentence-by-sentence translation path treated an unusable model reply as a fatal, unhandled
error: _coerce_translation_pairs raised a bare RuntimeError for a reply that was not JSON, that
had no pairs, or that was missing a translation for one of the units, and
_request_sentence_translation_batch_once did the same for an empty reply. The router's
except Exception turned that into HTTP 500 "Internal Server Error", and nothing was retried —
so one bad reply from the model ended the whole translation.

Free and smaller models produce exactly that kind of reply intermittently (prose around the JSON, a
truncated list, a dropped pair), which matches the "does not occur consistently" symptom in #427.
The JSON extraction was also too strict: it only stripped a code fence when the reply started
with , so a reply like `Here is the translation:\njson … ``` ` failed outright.

• backend/ai_logic.py
  • _extract_json_object replaces _strip_json_code_fence: it takes the JSON out of the reply
    regardless of surrounding prose (fenced block first, otherwise the outermost {…} span), and
    returns the text unchanged when there is no object at all — so a reply with no JSON is still
    reported as unusable instead of being silently accepted.
  • _UnusableSentenceReplyError is introduced as the internal signal for "this reply cannot be
    read as pairs" (invalid JSON / no pairs / incomplete pairs / empty reply).
  • _request_sentence_translation_batch_from_provider now handles that signal the same way it
    already handles context_length_exceeded: the batch is split in half and retried recursively.
    A single unit that still cannot be read is surfaced as
    ProviderRequestError(502, "The AI provider returned a sentence translation that could not be read. Please try again, or choose a different model."), so the user gets an actionable message
    and a meaningful status instead of a bare 500.

What I did not change

• The 422 "AI translation requires a configured provider API key." in the report comes from the
  API-key resolution path (_get_ai_api_key_for_provider). I could not reproduce it — that needs the
  reporter's credential-store state, and I have no OpenRouter key — so this PR does not touch it and
  is not a fix for that specific error message. It hardens the reply-handling path of the same
  translation flow.
• The single-shot /translate mode still raises a plain RuntimeError (HTTP 500) on an empty
  reply. That is a separate change; I left it alone to keep this diff focused.

How I validated this

Same test selection on both sides, no piped exit codes:

• Before the change (source change stashed, new tests in place): 6 failed, 16 passed — the six new
  tests are red with RuntimeError: AI provider returned invalid sentence translation JSON.
• After the change: 125 passed, 24 subtests passed across test_provider_request_retry.py,
  test_translation_units.py, test_ai_automation_logic.py, test_ai_provider_config_logic.py,
  test_ai_chat_endpoint.py, test_ai_models_env_api_key.py.
• Mutation controls: removing the tolerant JSON extraction turns only
  test_prose_wrapped_json_reply_is_accepted red; removing the unusable-reply branch turns the five
  recovery/error tests red while the parser test stays green — the tests measure the claimed
  behaviour.
• ruff check (0.15.10, the pinned version): All checks passed!; ruff format --check: already
  formatted; git diff --check: clean.

New tests in tests/test_provider_request_retry.py::TestUnusableSentenceReplyRecovery cover:
prose-wrapped JSON is accepted; an unreadable reply splits the batch; a truncated reply splits the
batch; an incomplete pair list splits the batch; a single unit that stays unreadable raises the 502
provider error; and the batch endpoint returns 502 with the readable message (no stubs, the real
router is exercised).

Not verified: the full suite. This machine has no uv, and the repository's uv sync --extra dev
pulls heavy optional dependencies (docling, markitdown[all], weasyprint, PyMuPDF, python-docx,
python-dotenv, markdown2), so I ran a minimal Python 3.13 environment. Running every module that
environment can import gives 8 failed, 188 passed, 13 collection errors (missing modules); the
identical selection on the unmodified baseline gives the same 8 failures, so they come from the
environment, not from this change. CI (ubuntu, Python 3.12, python -m pytest) covers the full suite.

AI assistance: this change was prepared with an AI coding assistant.

[petertzy]

Approved. The branch has been merged with the latest main, and the implementation addresses the intermittent sentence-translation failures caused by unusable provider replies.

Invalid JSON, prose-wrapped JSON, empty responses, truncated pair lists, and incomplete translations are now classified as retryable sentence-batch failures. The batch is recursively split so valid smaller requests can still complete. If a single unit remains unreadable, the API returns an actionable 502 response instead of an unhandled 500.

The JSON extraction accepts fenced or prose-wrapped JSON while preserving failure behavior for replies that contain no usable object. Regression coverage verifies successful parsing, batch splitting, single-unit error mapping, and the router response.

Validation passed: 125 translation-related tests, 497 Python tests, 157 frontend tests, frontend lint, Ruff check, and Ruff format check. Ready to merge.