What

Follow-up to the H5 tx_data retention work: lands the tx_data_memo seam + acceptance test + the work_source.cpp rewire as a distinct, additive PR off master (kept separate from #123, which stays segwit-SSOT-only).

build_connection_coinbase rebuilt the per-job tx-hex vector on every call, even when the mempool tx set was unchanged — the dominant churn/leak site in the 2026-06-02 heaptrack (work_source.cpp:634, ~768 MB across 1.2M calls).

How

• src/impl/btc/stratum/tx_data_memo.hpp — standalone, unit-testable seam. Memoizes the shared_ptr in a single slot keyed to a fingerprint over the merkle leaf set (wd->m_hashes, in leaf order). A repeat call against the same tx set returns the cached shared_ptr (a refcount bump) instead of re-serializing the whole mempool. The key is the exact leaf set that built this job merkle, so a hit is atomic-with-merkle by construction (no stale-tx-vs-fresh-merkle mismatch on submit). Single slot ⇒ bounded memory (one tx set retained), self-evicting on tx-set roll.
• work_source.cpp — rewires the a1 block to call the seam; guards the two new memo members (tx_data_fp_, tx_data_memo_) with template_mutex_ since the builder runs on connection threads (the call site does not otherwise hold it).
• work_source.hpp — adds the two memo members next to template_mutex_.
• src/impl/btc/test/tx_data_memo_test.cpp — standalone acceptance harness (style of template_parity_test.cpp, no btc/test CMake plumbing): proves pointer-identity reuse, correct invalidation on tx-set change, content correctness, and O(1) per-call cost on the unchanged path.

Scope / posture

• BTC-only; no LTC/DGB/DASH branch (per-coin isolation), no behavioral change to other coins.
• Signed commit, additive.
• NOT built — staged only under the capacity-return hold.
• PR-not-merge: merge stays operator-gated. Do not merge.