Make provider import state converge safely

[luca-ctx]

Summary

• persist convergent provider import outcomes and parser revisions
• fence inventory, completion, and missing-path transitions with a durable per-source generation
• keep transient I/O failures retryable while deterministic record rejection converges
• repair legacy failed rows and completed rows whose materialized session is missing

Why

Provider import state could previously remain ambiguous after mixed rejection, parser changes, or overlapping source scans. An older in-flight scan could also report against a newer observation. This change makes every durable result conditional on the exact current source observation without rewriting unchanged catalog rows.

Validation

• cargo test -p ctx-history-store catalog::tests (34 passed)
• cargo test -p ctx-history-store schema::tests (29 passed)
• cargo test -p ctx-history-capture codex (44 passed, 1 ignored manual benchmark)
• cargo test -p ctx-history-capture retained_existing_content
• cargo test -p ctx commands::import::native::tests (8 passed)
• cargo test -p ctx --test native_provider_rejections (16 passed)
• cargo clippy -p ctx-history-store -p ctx-history-capture -p ctx --all-targets -- -D warnings
• cargo fmt --all -- --check
• two independent adversarial review passes; all findings resolved or assigned to the separate write-pressure PR as preexisting scope

This is the first PR in a stacked import-indexing resilience series. It is ready for CI and review, but will be held from merge until the combined stack passes cross-platform and real-corpus validation.