fix: pay and finalize multi-batch merkle uploads (antd >= 0.12.0)

[Nic-dorman]

Fixes multi-batch merkle uploads against antd >= 0.12.0 (V2-1056), plus a latent allowance-formula bug the e2e exposed.

Linear: https://linear.app/autonominetwork/issue/V2-1056/multi-batch-merkle-uploads-fail-merklebatches

What

antd >= 0.12.0 splits uploads larger than one merkle tree (256 fresh chunks ≈ 1 GiB) into multiple MerkleBatches and leaves the legacy mirror fields empty, so the worker previously submitted an empty commitments list and failed opaquely.

• Batch planning (merkleBatchPlan): MerkleBatches when present (antd >= 0.12.0 always populates it for merkle, len >= 1), legacy single-batch synthesis as the pre-0.12 fallback, and a clear "needs antd >= 0.12.0" error when both are empty. All batches are validated (depth 1-8, 16 candidates/pool, parseable amounts) before any money moves.
• Payment loop (payMerkleBatches): one payForMerkleTree tx per batch (no ABI change — payForMerkleTree2() in antd docs is a naming slip for PaymentVaultV2's function, confirmed against evmlib + ADR-0003). Each confirmed batch is recorded before the next is paid (V2-426): one transactions row per batch, tx_hash = that batch's winner pool hash.
• Single approve tx (Signer.EnsureMerkleAllowance): pre-approves the whole plan's worst case, so the per-batch approval short-circuits (1 approve instead of N).
• Finalize: FinalizeMerkleUploadMulti with the index-aligned winner list when > 1 batch; the single-batch path keeps the legacy FinalizeMerkleUpload call unchanged.
• Salvage on mid-plan failure: a definitive payment failure (revert/RPC error) after >= 1 paid batch best-effort-finalizes with "" for unpaid batches — the paid batches' chunks store (surfaced as PARTIAL_UPLOAD), then the upload fails preserve-paid (errPaidNoRetry). A confirmation timeout never salvages: the tx may still mine, and skipping keeps the manual full finalize possible (the error carries the pending tx hash + batch index). If salvage unexpectedly fully succeeds (unpaid batches deduped), the upload completes normally.
• Allowance/cost formula fix (second commit, found live): the old bound — "sum over pools of max candidate" — is V1-era. PaymentVaultV2 charges median16(winner pool quotes) * 2^depth and picks the winner at execution time (PaymentVaultV2.sol:140), so batch 1 of the e2e reverted with ERC20InsufficientAllowance. Both the allowance bound and the max_gas_fee estimate now use the contract-accurate worst case (highest pool median << depth, summed across batches). The single-batch path had the same latent under-approval bug.

E2E proof (local devnet, 2026-08-24)

24-node devnet (merkle pools need 16 candidate peers; the 10-node preset can't quote) + anvil + antd v0.12.0, this branch's indelible:

1. Happy path: 1.15 GB random file → plan batches=2 (depth 8 + depth 5) → one approve tx → batch 1/2 paid 3.0 ANT, batch 2/2 paid 0.75 ANT (distinct winner hashes) → FinalizeMerkleUploadMulti → completed, 291 chunks, downloaded byte-identical (cmp clean). DB: two transactions rows with the exact per-batch winner hashes/amounts; actual_cost = 3.75 ANT = sum.
2. Failure drill: fresh wallet funded with 3.2 ANT (batch 1 needs 3.0) → batch 1 paid, batch 2 reverted (insufficient balance, definitive) → salvage finalize stored chunks_stored=256, chunks_failed=35, total_chunks=291 → upload failed/paid_unfinalized, temp source preserved, exactly one transaction row.
3. Retry pays only the remainder: re-upload of the same content with a funded wallet dedups the 256 salvaged chunks and pays only the ~35-chunk remainder.

Unit: go test ./... green (batch planning/fallback/version-gap error, validation, pad/salvage decision, contract-formula payout bounds, classifyFailure rows for the new wrapping shapes).

Notes for review

• merkleAllowancePayer is an optional interface seam so the uncommitted hosted-payment work (payer interface) merges without changes — payers lacking the method fall back to per-batch self-approval inside PayForMerkleTree.
• Pre-existing, out of scope: RequeueStuck(60) headroom vs N×5-min confirmation waits on very large files — follow-up ticket suggested.

🤖 Generated with Claude Code

[dirvine]

Reviewed at exact head 85228779cc6701945fa57388636ff053e8334a2c.

The multi-batch flow is consistent with antd-go v0.12.0 and PaymentVaultV2: the allowance bound uses the upper median quote multiplied by 2^depth, payments are nonce-serialised, and the timeout/salvage branches avoid automatically re-paying an unconfirmed transaction. The single-batch compatibility path and multi-finalize ordering also match the SDK contract.

Verified locally:

• go test ./... -count=1
• go test -race ./... -count=1
• go vet ./...
• frontend npm ci && npm run build
• git diff --check

Required exact-head CI is green (Docker, frontend, lint, smoke, PostgreSQL and SQLite).

Non-blocking follow-up: validateMerkleBatches does not yet fully enforce its “reject malformed data before any money moves” comment: it could also validate pool-hash syntax, the contract-required pool count, and non-negative uint256 candidate amounts. The later signer/contract checks plus partial-finalize path keep this from being a merge blocker, but validating those fields up front would make the stated invariant complete.