v0.32.5

Fixes

One long message could freeze segmentation permanently (#320)

Text pulled from multi-part messages — any assistant reply containing a tool call, any user message with an attachment — was never length-capped when building embedding windows. A single long message pushed its window past the embedding model's 8192-token limit, which 400s the entire batch. indexSession then threw before advancing segment_state, so the same messages were retried forever: no new summaries, and under the 0.32.3 trim clamp, no trimming either. POST /segments/start sits behind the same code path, so the recovery path was inside the broken thing.

Boundary-detection text is now capped, and a rejected embedding batch is retried value-by-value with shrinking input, so one pathological window degrades instead of stalling the index for good.

Affected agents unblock on upgrade plus a single POST /segments/start. Truncation is read-time only — no schema change, no migration, stored content untouched.