[nicoloboschi]

Fixes #5715.

Problem

We size TypeSafe requests with our own tokenizer, but Jev counts mixed-script text (emoji, CJK, Arabic, accents) at about 2× ours. A request that fits our 26k budget can come back 400 {"detail":{"error_type":"max_tokens_exceeded"}}:

• TypeSafe reranker → recall fails with HTTP 500 (the issue).
• Decision model ask_all → the failure was logged and those answers silently left out; the decision reranker then ranked those candidates last.

Jev's tokenizer isn't published and there is no count endpoint, so we can't count exactly. Instead, recover when Jev refuses.

Fix

• TypeSafe reranker (cross_encoder.py): rank, cut, evidence sufficiency and page impact rebuild an overflowing request at half the budget (26k → 13k → 6.5k → 3.25k), then raise.
• Decision model (decision_model.py): ask() splits an overflowing request into halves until a single question is too long; ask_all() now raises instead of dropping answers.
• Decision reranker: caps each memory at 4k tokens, so one long memory can't overflow its question and fail the whole rerank.

No long-context header and no new setting, unlike the patch proposed in the issue: that was opt-in, so the default would have stayed broken.

Verified

• Against the live API with the issue's repro (7 mixed-script candidates, 19k tokens by our count): the rank and cut each overflowed once, retried at 13k, and returned scores.
• Tests: the fake upstream refuses bodies over a window counted at 2× ours: reranker repack, evidence retry, give-up after the floor, decision-model halving, ask_all raising, memory cap.