fix: remap compaction references when forking sessions - #1420

Fixes #1419.

Forking gives transcript records new UUIDs while compaction metadata still points to the source IDs. This changes the shared disk/store fork builder so copied messages and their known compaction references use the same mapping.

• Remap preservedSegment.headUuid/tailUuid/anchorUuid, preservedMessages.anchorUuid/uuids/allUuids, and runningTurnSends.openerUuid/uuids/summarisedOnly.
• Generate IDs only for records actually written. Progress records remain available for parent traversal.
• For partial forks and sidechain/progress exclusions, trim known excluded IDs from lists. Drop a metadata block when its anchor/opener or segment endpoint was excluded; do not pull excluded messages into the fork or guess replacement endpoints.
• A fork whose cutoff is a compact boundary ends just before that boundary. The boundary's preserved blocks are anchored after it, and @tonydzi measured that resuming a fork that kept the boundary without them lost the earlier compaction summaries (CLI 2.1.317 against a stub).
• Preserve IDs already absent from the source, malformed values, and uninterpreted fields in retained blocks. Historical running-turn bookkeeping can outlive transcript records, so a blanket UUID-shaped-string validator would reject legitimate input.
• Copy changed metadata containers; source entries and arbitrary message content remain unchanged.

Retains the original parameterized disk/store, segment/list/both, partial-fork, malformed-metadata, and source-preservation coverage. Adds 56 cases across both public APIs and both AnyIO backends, with schema-aware assertions against the IDs actually emitted. These assertions are test safeguards, not runtime rejection or automatic discovery of future metadata fields.

Validation on main 2d485201617258867dd0f74b9e7307056f5938fe plus this branch (Python 3.12, MCP 2.3.0):

• Affected three test files: 215 passed.
• Full tests/ suite: 1,701 passed, 6 skipped. The earlier timing assertion failure did not recur; no timing code or tests changed.
• Ruff check, formatting, mypy and git diff --check: pass.
• New regression module against isolated current-main snapshots: 56 failed on main; 44 failed, 12 passed with either the original fix: remap compaction references when forking sessions #1420 or fix(sessions): remap compactMetadata preserved message UUIDs on fork (#1419) #1428 fork implementation substituted. These are targeted comparisons, not full-suite runs of the other PRs.

Update for the boundary cutoff (e70a802, Python 3.12):

• Full tests/ suite: 1,713 passed, 6 skipped. The three new boundary-cut tests (both APIs, both AnyIO backends) fail on the previous head and pass now. The existing store partial-fork test now expects the fork to stop before the boundary.
• Ruff check, formatting and mypy pass on the changed files.

The new evidence is synthetic public-API testing and local CLI code inspection. No end-to-end CLI/model resume was run; existing malformed or incomplete source metadata is not repaired.

Thanks to @tonydzi for the release reproduction and running-turn field report, @xg-gh-25 for the partial-fork/integrity suggestions, and @somuai for the overlapping implementation in #1428. This retains #1420's boundary-scoped helper and broader regression coverage; the follow-up can move to whichever PR maintainers select so only one implementation lands.

Developed with Codex assistance.

Co-authored-by: Soumyajit Ghosh [email protected]