Fix truncated translate-and-adapt; route through shared Haiku client

[mircealungu]

Problem

A user tried to import a fanfiction chapter (The Changing of the Guard, ~4k words / 24k chars) via translate-and-adapt (English → German B2) and it failed. Server logs showed the Anthropic call returning Error parsing Anthropic JSON: Unterminated string — the LLM reply was cut off at max_tokens mid-JSON, then fell back to DeepSeek (also capped) and produced a broken/incomplete article.

Root cause: not a length limit and not abuse — the output caps (max_tokens=4000 Anthropic, 8000 DeepSeek) were too small for a full chapter's translation, which runs at least as long as the input. There was no truncation detection, so a cut-off reply became unparseable JSON.

Changes

• Raise the Anthropic cap to 16000 (Haiku 4.5's output ceiling is 64k) and route the call through the shared haiku_client instead of a duplicated hand-rolled requests.post — model (models.SIMPLIFICATION), key, endpoint, and truncation handling now live in one place. Removes ~40 lines of duplication.
• Detect truncation in haiku_completion (stop_reason == "max_tokens" → None, fail-soft) so a cut-off reply becomes a clean fallback/failure instead of garbage JSON. Left haiku_completion_or_raise (the batch crawl path) unchanged on purpose — long crawled articles can legitimately hit the cap there, and raising would shrink simplified-article inventory.
• DeepSeek: raise to its 8192 ceiling (16000 would 400) and detect finish_reason == "length".
• Endpoint: return 422 with an actionable message ("try importing a single chapter") instead of an opaque 500 "Translation failed".

The reported chapter should now translate on the primary Anthropic path (16k leaves ample room). The truncation detection is the safety net for genuinely huge single chapters.

Not included / follow-up

• Chunking long works into per-chapter LLM calls is the durable fix for arbitrarily long content — tracked separately.
• A weekly per-user character cooldown (the original idea) is not needed: this was a truncation bug on a normal-sized chapter, not a cost/abuse problem. It pairs naturally with chunking later, if wanted.

Testing

Import + wiring smoke-tested (HAIKU_MODEL resolves via models.ANTHROPIC_HAIKU). No automated tests cover these methods (they call live LLM APIs). Recommend re-importing the fanfiction URL to confirm end-to-end.

🤖 Generated with Claude Code

[github-actions]

ArchLens - No architecturally relevant changes to the existing views