feat: ironwood support

[dorianvp]

This branch makes the wallet ironwood-capable from the network edge to the command line. The pepper-sync crate scans the third shielded pool — ironwood compact-block actions, a third shard tree, and a third nullifier map — counts it in sync progress and batching budgets, and persists it behind version-gated wallet-format bumps (SyncState v4, NullifierMap v2, TreeBounds, WalletTransaction, and ShardTrees v1), each with migration tests over old blobs. Following ZIP 326, receivers and viewing keys are scoped to the Orchard protocol rather than to a pool, so ironwood shares orchard key material and unified addresses gain no new receiver. On top of that, zingolib extends note selection, balances, summaries, and the zcash_client_backend trait implementations to the new pool, riding the ironwood dependency train (zcash_client_backend 0.24.0-rc.1, zcash_primitives 0.29.0, orchard 0.15.0) plus a [patch.crates-io] git pin of lightwallet-protocol that carries the ironwood proto fields until upstream regenerates and publishes them.

The largest single addition is the consent-bound orchard-to-ironwood migration engine of ZIP 318, in the new zingolib::wallet::migration module: deterministic denomination sizing and note splitting (split), part records with an explicit state machine (parts), bucket-scheduled broadcasts with jitter (schedule), versioned persistence (store), restart reconciliation (reconcile), and an immediate drain path (drain). LightClient methods orchestrate consent, rounds, catch-up, and status reporting, and zingo-cli exposes the corresponding migrate commands, with shield now targeting the ironwood pool. This work arrived through PRs #2428 and #2429, the summary and value-transfer extensions of #2463 and #2466, the sweep-floor fix of #2468, the zebrad-testing line of #2461, the stable backport of #2473, and #2475, which also carried #2477's scan-progress accounting fix and its regression test.

The validation stack behind the branch runs zebrad 6.0.0 and zainod 0.6.0-rc.1-no-tls inside a content-addressed CI image (#2472, #2476). The tip_spend_rejection suite documents a completed investigation: the boundary-adjacent orchard-output rejection that motivated three rounds of hypothesis discrimination was a zebra 6.0.0-rc.0 mempool-admission bug, fixed upstream, and the suite now pins acceptance as a regression sentinel. Because zainod 0.6.0 masks validator rejection text (zingolabs/zaino#1404, which holds the indexer at the release candidate), test failures are classified by a new attribution instrument in zingolib_testutils that judges the wallet's retained transaction bytes directly against zebrad's JSON-RPC, twice — at failure time and after five blocks of distance — and names the responsible layer.

Three known reds remain on this branch and are recorded rather than resolved: mine_to_ironwood, send_mined_ironwood_to_ironwood, and send_orchard_back_and_forth fail deterministically on freshly mined zebrad 6.0.0 chains, each observing a balance exactly one post-funding-stream block reward (618,780,000 zatoshis) below expectation; they pass locally only where chain caches mined under rc.0 replay, because cache manifests key on the validator's type rather than its version. PRs #2469 (OP_RETURN spend pipeline) and #2470 (Nym transport design) remain open against this branch. The table below classifies every changed line by package and by logic type, using file paths as the classifier.

package	domain	test	tool	docs	design/plans
(workspace root)	—	—	+118/−53	+135/−3	+245/−0
libtonode-tests	—	+1045/−226	—	—	—
pepper-sync	+2128/−440	+6/−1	—	+2/−0	—
zingo-cli	+263/−5	—	+1/−0	—	—
zingolib	+7088/−628	+504/−196	+2/−0	+17/−7	—
zingolib_testutils	—	+222/−22	—	—	—
total	+9479/−1073	+1777/−445	+121/−53	+154/−10	+245/−0

[zancas]

The audit is complete. All five putative high-severity issues of #2493 were real — each harm test fails on its fix commit's parent — and all five are genuinely fixed on this branch, where the same tests pass. The full verdict, test by test:

1. Stranding. Pre-fix, start_ironwood_migration accepted a fragmented wallet's consent and persisted a Planned state nothing could advance; on HEAD it refuses, persists nothing, and drain_orchard_to_ironwood is never told AlreadyInProgress.
2. Pass abort. Pre-fix, one spent bound note hard-errored the whole broadcast pass (MigrationBuild("bound note is already spent…")) and a healthy signed part in the same window went unsent; on HEAD the bad part skips to reconciliation and the healthy part's bytes reach the broadcast endpoint in that same pass.
3. Consent collapse. Pre-fix, migrate_to_ironwood over a consented schedule went straight for the network (Err(Offline) on an offline client) before any consent re-check; on HEAD it refuses as a migration-domain decision before any network action, and every part keeps its bucket.
4. Rerun no-op. Pre-fix, a Complete state's stale confirmed parts survived entry and stood ready to defeat binding — the root of the fee-burning empty-part_txids success; on HEAD terminal history clears so a rerun binds fresh funds.
5. Mobileclient drift. Pre-fix, unfunded_mobileclient bound to the workspace-default schedule (the silent no-diff-line drift); on HEAD a source-walk test pins every *_mobileclient body away from default_test_activation_heights() and pins the root to the deferred schedule, so the drift class cannot recur silently.