Investigate Tamil generation/translation quality

[praneethaneelapareddigari]

Several Tamil outputs in Pilot v1 were repetitive/incoherent independent of the refusal question (§8.2). Use the translation QA / back-translation gate (methodology §5.1) to determine whether this is a translation-side or generation-side issue before trusting any Tamil result.

[praneethaneelapareddigari]

Update: Root cause found — this was two separate problems

See docs/tamil_quality_findings.md for full analysis.

Problem A (fixed): translator.py used a fixed max_new_tokens=256, which
truncated longer/denser prompts for Chinese, Tamil, and Malay before translation
completed — while Hindi consistently fit within budget. Confirmed with a literal
mid-sentence cutoff in the Chinese translation of mis-002 (ends with a trailing
comma, the actual instruction never translated), and adv-002 losing its entire
harmful request in Chinese/Tamil/Malay simultaneously (33% of original content
survived). Fixed by scaling max_new_tokens with input length instead of a
fixed cap, plus an automated completeness check that now flags any suspiciously
short back-translation for review.

Problem B (confirmed, separate, not fixable in code): even with fully intact,
correctly-translated Tamil prompts (e.g. cyb-002, rol-002), mistral's Tamil
generation was still repetitive/incoherent. This points to a genuine model
capability limitation in these 7B-class open models, not a pipeline bug — worth
factoring into model selection for the 300-500 prompt study.

Next step: re-run translation with the fix and confirm all back-translation
ratios land near 1.0 before scaling the dataset (Issue #4).