fix: keep exact common ancestor when checking out to the past

[lodystage]

Summary

Fixes #1056 — forkAt(frontiers) hanging the browser with no error on large documents.

_find_common_ancestor discarded the ancestor it had just computed and returned empty
frontiers whenever a branch of its DAG walk reached a root without meeting the other
side. Checking out to a version already in the past — exactly what forkAt does —
always ends that way:

find_common_ancestor([1:10, 2:5], [1:10]) -> Frontiers([])   // should be [1:10]

The conservative fallback is now skipped when the computed ancestor is exactly the
target version.

Rationale

An empty replay base is not a constant-factor cost, it is document-sized and
multiplicative: the causal iterator walks the entire oplog, so every container in
the document gets a diff calculator, and each of those independently re-walks the whole
oplog to rebuild its tracker — O(containers x ops).

When ans == right, every id of right is reachable from left, so right is the
maximal common ancestor and there is nothing to be conservative about. It is also the
safe direction for the trackers: from_vv then strictly contains to_vv, so the
retreat is never empty and the richtext/list/movable-list trackers are still rebuilt
from CRDT ids rather than advanced incrementally.

The symmetric relaxation (ans == left, fast-forwarding past a concurrent branch) is
not safe — it has no retreat to force a rebuild, and undo_tree in crates/fuzz
fails under it. The code comment names that test so the next person does not retry it.

The regression came from b81abfce (#974) in 1.13.7, which is why 1.13.6 is unaffected.

Validation

• Full gate cargo test --features=test_utils,jsonpath --no-fail-fast: 73 groups,
  1417 tests, 0 failures

• crates/fuzz 128 tests including the undo_tree canary

• New common_ancestor_of_subset_target_is_the_target_itself (dag unit test) and
  test_fork_at_with_concurrent_frontiers (fork / checkout / attach / re-import)

• Perf harness crates/examples/examples/fork_at_many_containers.rs, 1600 containers /
  194k ops:

  	1.13.6	1.13.7	this PR
  checkout	46.7ms	268ms	1.45ms
  forkAt	193ms	679ms	346ms

Known limitation

forkAt stays above 1.13.6 because it internally checks back to the original version,
which is the ans == left direction that cannot be relaxed safely. Separately, a single
large text container is still ~2x slower than 1.12 due to the unconditional tracker
rebuild b81abfce added. Both are recorded in context/CONTEXT-GAPS.md, along with
four relaxations of that rebuild that were measured and ruled out — including the trap
that an oracle checking only a single checkout wrongly reports the rebuild as redundant.

Background and the cost model live in the new context/diff-calc-replay-base.md.

🤖 Generated with Claude Code

[lodystage]

Superseded by #1058, which reworked replay-base selection around critical versions — all_tips_covered_by_ancestors distinguishes a genuine concurrent branch from a redundant path into a known ancestor, and latest_single_head_critical_version retreats to the latest safe cut instead of to empty frontiers. That covers the #1056 hang this branch targeted.

The argument here was also unsound, not just redundant: it treated the meet of the two versions as a safe replay base whenever the target was already in the source's past. For current = [10@1, 5@2] with peer 2 branching at 3@1, the meet 10@1 is not a critical version, since 0@2..5@2 is concurrent with 4@1..10@1. It passed the fuzz corpus by luck rather than by correctness.

Closing rather than reworking. Nothing here is needed on top of #1058.

[github-actions]

WASM Size Report

• Original size: 3104.95 KB
• Gzipped size: 1031.21 KB
• Brotli size: 722.44 KB