Fix critical resolution and recovery defects; prepare v0.2.0

[script3r]

This completes the persistent correctness audit and prepares the first 0.2.0 release. Accepted records always finish entity resolution before commit; rejected batches cannot leak staged records into later requests.

• Enforce durable record ownership, rollback partial ingest state, and preserve interner IDs through bounded-cache recovery.
• Guard transitive cross-shard components with all temporal strong-ID observations and assemble complete canonical entities before query conjunction and golden mastering.
• Keep common keys and memory-saver matching correct; persist complete tiered index buckets through eviction and updates.
• Reuse authoritative membership for selective queries, dispatch shard requests concurrently, and lend staged records to parallel extraction.
• Repair Rust 1.98 CI lints, validate every deployment shell script and release package, and gate publication on complete reusable CI.

Live protocol 6 requires a coordinated router/shard/replica upgrade. Existing WAL/checkpoint version 1 and durable reservation format 5 remain compatible. Composite query labels preserve global collision semantics.

The release includes enabled persistent regressions for all eight audit defects and follow-up recovery/protocol/query edge cases. Full validation and same-machine performance measurements are recorded in docs/critical-audit-2026-09-05.md; only the manual timing diagnostic is ignored.

Validation: 356 local tests passed (three intentional ignored diagnostics/doctests); Rust 1.98 Clippy, Rust 1.88 MSRV, formatting, actionlint, and package verification passed. Five persistent shards acknowledged all 1,000,000 load-test records with zero failed batches; throughput was 43.4K before and 43.9K after. The 80K-record selective query after unrelated ingest fell from 300 ms to 0.008 ms. GitHub CI passed all three jobs, including Linux tests/lints/package/manifests, MSRV, and the production container.