fix(cache): attempt the companion delta BEFORE the block table is trimmed
The MLLM delta derive landed inert. It ran after trim_block_table, so it reconstructed attention KV at the block-aligned floor while the companion it had to pair with sat above that floor, and its own offset guard then correctly refused the seed. Measured on a growing Qwen3.8 chat: checkpoint 38467, trim to 38464, full hit 38528 -- and zero DELTA lines in the log, with the hit still discarded for a full prefill. Only the untrimmed hit can be sliced to the checkpoint. The attempt now runs first, and the trim is reached only once the delta has declined, so the existing resume-or-abandon path is unchanged when it does. Caught by running the reproduction rather than trusting the unit tests, which exercised the helper directly and so never saw the ordering.
J
Jinho Jang committed
e89006fb387009cb1c37a71b83ecb24012c35614
Parent: 0655d33