Fix March 2026 production panic regressions

[zxch3n]

Summary

• fix stale DAG causal iteration after lazy node splits so export/import no longer drop same-peer ranges
• fix richtext IdToCursor fragment-tail updates and zero-width update handling
• add a list import_batch() regression that exercises the same split-tracking path seen in production
• recover document-owned locks after poisoning instead of cascading PoisonError panics
• add unpoisoned lock helpers across core internals so ordinary Mutex users follow the same recovery rule

Verification

• cargo test -p loro-internal stale_iterator_state_repairs_missing_same_peer_continuation -- --nocapture
• cargo test -p loro-internal larger_node_only_advances_consumed_slice -- --nocapture
• cargo test -p loro-internal repeated_tail_splits_keep_id_to_cursor_consistent -- --nocapture
• cargo test -p loro-internal zero_width_small_update_keeps_insert_set_non_empty -- --nocapture
• cargo test -p loro-internal large_update_can_replace_the_tail_range -- --nocapture
• cargo test -p loro-internal list_import_batch_stays_consistent_after_repeated_tail_splits -- --nocapture
• cargo test -p loro-internal poison -- --nocapture
• cargo check -p loro-internal

Notes

• the active ImVersionVector export panic still was not reproducible in this checkout; no speculative fix was included for that path
• the FFI-side unpoison work lives in /Users/zxch3n/Code/loro-ffi on branch codex/unpoison-locks and is not part of this PR

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a28c207df2

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Use sync lock types before calling unpoison helpers

read_unpoisoned() here is being called on std::sync::RwLock, but RwLockExt is implemented only for crate::sync::RwLock in sync.rs (same pattern for MutexExt). That works in non-loom builds only because crate::sync aliases std there; under cfg(loom) the alias switches to loom::sync, so this method no longer exists on std::sync::RwLock and the loom configuration fails to compile. The same regression pattern appears in other files that still import std locks directly (e.g. awareness.rs, pre_commit.rs), so these lock fields/imports should be switched to crate::sync::{Mutex,RwLock} (or the trait impls broadened) before using the helper methods.

Useful? React with 👍 / 👎.