2026-10-22 — Owner-requested initial/manual full recovery only
Valid per-account discovery hints no longer expire. Elapsed time, clock rollback
and lower chain tips do not force full change recovery. Initial synchronization
(no valid hints) and explicit Full recovery scan still search change0..1999.
Quick refresh continues to read current balances/history for the remembered
prefix plus gap20. New distant change requires manual full recovery. The older
daily/full recovery conditions below describe previous releases only.

2026-10-21 — 0.1.17-preview connection lifecycle rollback
The 0.1.15 release source archive and 0.1.16 commit were compared. Their scan
implementation after scan(seed, ...) is identical, including daily/full recovery
conditions, persisted discovery keys and change indices 0..1999. The 0.1.16
change shared the receipt-monitor connection with balance/send operations.
0.1.17 restores a dedicated monitoring socket and fresh lookup/send connections.
Monitor retries or unmounts cannot close the lookup connection. Public cache,
lazy ownership, batched history and 60-second monitoring reconciliation remain.
Two connections per open wallet may be present; connection-count capacity must
not be inferred from the former one-socket design. No server setting changed.
Controlled PC tests of each released scanner used a 20-second test deadline:
neither full scan finished under the current endpoint/path conditions. These
sequential observations are not a reliable version speed comparison. A small
8-history-query batch took ~2ms on server loopback and ~507ms from this PC.
These measurements do not reproduce the S23 UI issue or prove its root cause.
No address count cap or new global 60-second lookup limit is introduced.

2026-10-20 — 0.1.16-preview monitoring update
Empty monitoring startup derives 20 receive + 20 change addresses, instead of
4,000 ownership keys. Reconnects reuse a public-only cache scoped to one open
wallet. Restored wallets extend the window using validated discovery metadata.
Actual new-receipt classification may still derive all supported ownership keys
once; this protects high-index change/self-transfers and does not run on every
empty startup. Such first classification can still be slower on a phone.
History reads use bounded batches; periodic reconciliation is 60s and live pushes
remain immediate. Monitoring, balance and sending share one verified TLS socket.
Public endpoint test from this PC: 1 connection, 77ms handshake, 2046ms initial
monitor result, 40 derivations, 26 frames / 43 logical requests. No broadcast or
private account used. This is not a physical-device SLA or a capacity benchmark.

2026-10-19 — 0.1.13-preview address discovery correction

Same local isolated Core/ElectrumX lab, not a physical phone or Internet SLA.
Historical baseline: 140664ms, 4044 logical RPC requests for an empty full scan.
Updated cold full recovery: 3984ms, 4044 RPC requests in 131 wire request frames.
New client/connection with persisted public hints: 3038ms, 84 RPC requests in
8 frames. Forced full recovery: 5020ms. Small-account latency is still around
3 seconds here; the RPC reduction does not promise instant balance refresh.
Batching reduces round trips, not the full recovery's logical server workload.
All ten integration checks passed again, including real signed transactions,
mempool notifications, confirmations, reorgs, QR-PSBT, expiry and index restart.
Local server notification latency was 1.666 seconds, not phone display latency.
Machine-readable evidence: discovery-performance-results.json.
Historical evidence: BACKEND-VALIDATION.txt and backend-lab-results.json.

Design
- At most 32 correlated read RPCs per batch; replies can arrive out of order.
  Only explicit unsupported-batch errors permit individual-read fallback.
- Public per-account used indices and last full-scan time/height survive restart.
  Invalid/foreign hints are rejected. Cached balances never drive spending.
- Normal refresh rechecks the allocated prefix plus gap20 on both branches.
  A high historically used change index still means a large normal query range.
- Initial synchronization (missing valid hints) or manual Full recovery scan
  searches all change indices 0..1999. No time/clock/lower-tip trigger. Receive gap20
  and cap2000 remain. Newly used distant change from another device requires a
  full scan; the normal window cannot discover arbitrary unused gaps.
- Raw transaction and derived-script memory caches avoid repeated decoding and
  derivation. Each refresh obtains current history/UTXOs and checks a stable tip.
  Source txid/script/value, maturity, expiry and fixed genesis checks remain.
- No secret format, derivation or QR transport migration. Existing change25/1999
  recovery regression tests retained. Six additional tests bring the total to53.

Still required: public VPS/domain/trusted TLS and production indexer review;
physical device cache/button/auth/camera/QR/backup checks; funded mainnet tests.
No pipeline or CI settings were accessed or changed.
