refactor(state): apply ironwood-migration polish from review of #1379

[zancas]

Fixes #1382.

This applies the six non-blocking polish items surfaced in review of #1379, in one commit:

1. Backfill errors carry remediation advice. When the v1.2.1 → v1.3.0 migration cannot refetch a post-NU6.3 block, the error now tells the operator to ensure the backing validator serves the range from NU6.3 activation through the database tip, or to wipe and re-index.
2. No per-height allocations in put_idempotent. The helper takes a lazy describe closure, so the per-row success path allocates nothing; the previous version formatted two messages per height.
3. put_idempotent hoisted to module scope and reused at all three sites that inlined the NO_OVERWRITE + verify-match pattern: the spent-batch flush, the txid_location rebuild, and the v1.3.0 commitment/ironwood writes. At the txid_location site the byte-for-byte comparison subsumes the old decode-and-checksum check, since a corrupt existing row cannot match the rebuilt bytes.
4. Activation heights resolved in one place. A new PoolActivationHeights struct in finalised_state.rs is shared by write_blocks_to_height and the migration, so the two call sites cannot drift apart.
5. Visibility narrowed. FinalisedSource::start_validator, both ironwood_db accessors, and DbV1::start_validator go from pub(crate) to pub(super) — the minimum that compiles.
6. multi_thread flavor justified in place. The regression test gains a comment explaining why it needs the multi-threaded runtime (the validator runs blocking LMDB validation inline on its task).

Verification: cargo check, cargo clippy --all-targets, and cargo fmt are clean (the one clippy warning is pre-existing in an untouched file), and all six finalised-state migration tests pass under cargo nextest run -p zaino-state migrations:: — including the crash-resume and txid_location-rebuild tests that exercise the shared put_idempotent at every call site.

🤖 Generated with Claude Code