fix(session): remap compaction tail_start_id when forking

[spark4862]

Issue for this PR

Closes #24700

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Fixes Session.fork not remapping CompactionPart.tail_start_id, which caused forked sessions to include all pre-compaction messages in context and inflate session length.

Problem

Session.fork copies messages and remaps each message’s id and assistant parentID via idMap, but tail_start_id on compaction parts was left as the parent session’s message id.

filterCompacted (message-v2.ts) uses tail_start_id as the retention boundary: it walks older messages until it finds one whose id equals tail_start_id. In a fork, every message id is new, so the old tail_start_id never matches; the loop does not stop where compaction intended, and pre-compaction history leaks into the prompt (observed as much larger context / context_length_exceeded on the next request).

Regression

Introduced when tail_start_id and the corresponding filterCompacted behavior were added in commit 6f5a3d30f (“keep recent turns during session compaction”, 2026-04-10). Session.fork was not updated to remap that field alongside parentID.

Fix

When cloning parts, after ...part, override tail_start_id for compaction parts using the same idMap:

...(part.type === "compaction" && part.tail_start_id
  ? { tail_start_id: idMap.get(part.tail_start_id) }
  : {}),

If the tail message was not copied (e.g. partial fork), idMap.get returns undefined, clearing the stale reference instead of pointing at a non-existent id.

How did you verify your code works?

• Automated: from packages/opencode, run
  bun test test/session/messages-pagination.test.ts -t "fork remaps compaction"
  (adds test fork remaps compaction tail_start_id for filterCompacted).
• Manual: reproducing Session.fork does not remap CompactionPart.tail_start_id, breaking filterCompacted and inflating forked context #24700 — forked session context no longer balloons (e.g. stays ~300k instead of jumping toward ~800k).

Screenshots / recordings

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[rekram1-node]

/review

[github-actions]

lgtm