fix(session): prune tool outputs before serializing compaction request

[YOMXXX]

Summary

Fixes #1661. Compaction fails with "context exceeds model limit" even when the thread's live token count is well under the model window (e.g. 100k used vs 1M window) — large tool outputs serialized into the compaction request inflate it past the limit.

Root cause: packages/opencode/src/session/compaction.ts has a DB prune() for tool outputs that is defined but never called. compaction.process builds its modelMessages from selected.head via toModelMessagesEffect WITHOUT pruning old tool outputs, so a session with big tool bodies (file reads, diffs) overflows the compaction model.

Fix: strip old tool outputs on the in-memory copy before serialization:

• New pruneToolOutputs(messages) mirrors the DB prune walk (protect last 2 user turns + summary boundary + PRUNE_PROTECTED_TOOLS, accumulate past PRUNE_PROTECT, strip output and providerOutput).
• Called in compaction.process after structuredClone(selected.head) and before toModelMessagesEffect — never touches the DB or the visible transcript.

Test Plan

• 4 focused tests: strips old outputs past PRUNE_PROTECT; keeps protected skill outputs; small sessions untouched; summary boundary stops the walk.
• bun typecheck clean; session suite 913 pass no regression.

Notes

• The in-memory strip is the right scope for Compaction fails - context exceeds model limit (it shouldn't) #1661 (it bounds the compaction request). The DB prune() remains unwired (separate storage-savings enhancement).
• Known limitation (documented in review): tool input and non-provider-executed structured output are not bounded — the fix targets the common large-output case (read/bash/edit/write).