Remove shared reader bottlenecks from WTI lookups

[pathscale]

Summary

• Publish an immutable chunked route index protected by a local ps-reclaim domain; structural writes copy the root vector and only touched chunks, not the whole BTreeMap.
• Use per-node RwLocks for concurrent point reads, with bounded generation retries and a canonical topology-lock fallback.
• Drop reclamation pins before blocking on node locks; retire roots after the even-generation release and amortize registry scans at a 64-root backlog.
• Add bulk attach_nodes/attach_multi_nodes reconstruction and replace the misleading mutable-looking iter_nodes API with snapshot_nodes -> Vec<Node>.
• Normalize duplicate search semantics across all search backends and preserve acknowledged inserts across stale topology repairs.

Review resolution

• B1: full-map publication clone and unbounded hot spin removed.
• B2: persisted reconstruction API can publish once per bulk load. The corresponding WorkTable caller migration is prepared in WorkTable PR #96 and must merge after this crate.
• B3: readers never retain a reclamation pin while blocking.
• B4: two retries followed by the canonical progress path; caller code runs after releasing the structural lock.
• B5: semantic trap removed through the explicit snapshot_nodes API.
• M1: writer-interference latency and throughput were validated with the separate local wt-bench harness; that harness is not part of this PR.
• M3/M4: recursive-lock invariant documented and duplicate search returns the first match for every backend.
• M5: the private retirement wrapper remains T: Send, Node: Send; it can only destroy a uniquely retired pointer and cannot expose or dereference it.
• M6: visibility/oracle tests, a Loom weak-memory model, and a manually triggered multi-seed Miri workflow cover publication and reclamation.
• M7: version policy intentionally unchanged for the coordinated beta release.

The follow-up audits also added a dedicated last-node rekey operation, stale-high/stale-low attachment repair, bounded canonical rebuild on bookkeeping divergence, split/merge hysteresis, miss-before-copy removal, and symmetric recovery tests.

Raw performance

131,072-key pure reads, threads 1/2/4/8 (M lookups/s):

• Arctic control: 13.644 / 26.843 / 53.410 / 107.62
• Candidate WTI: 4.1442 / 7.8045 / 15.218 / 28.462
• Released WTI 0.0.11: 3.7507 / 7.2218 / 6.5732 / 4.4209

At 8 readers the candidate is 6.44x released WTI and scales 6.87x from one reader.

Final local writer-interference run: one million monotonic structural inserts, four stable-key readers, three repetitions (medians):

• Candidate quiet: 19.626 M reads/s
• Candidate busy: 17.367 M reads/s (-11.5%)
• Busy p99: 459 ns; p99.9: 5.25 us
• Writer: 152.75 ns/insert

Bulk reconstruction after the audit fixes remains linear: 8,000 nodes in 1.072 ms.

Validation

• 127/127 all-feature library tests plus Loom on the final local commit.
• Loom publication-order model passes; weakening the final generation Release to Relaxed makes it fail.
• Reclamation interleaving passes Miri seeds 0..8 against local ps-reclaim; stale-boundary and recovery paths pass seeds 0..4.
• Strict all-target/all-feature Clippy, formatting, diff check, and packaged-crate compilation pass.
• Automatic, heavy, and manual Miri workflows run on the final commit before merge.