core+gui+cli: anticipation/rebuild dates, preserve relative settings

Fixes around the delivery/build/check date handling (karen7 regtest):

- build_will (GUI + CLI) re-anchors date_to_check to the CURRENT heirs'
  earliest future delivery before building, so an anticipated rebuild is no
  longer blocked by the stale old-will anchor (NO_FUTURE_DATE). The checks of
  the existing will keep their anchored date_to_check.
- _sync_locktime_to_built_txs now PRESERVES RELATIVE locktime/threshold
  recipes ("2y"/"150d") in WILL_SETTINGS instead of freezing them to absolute
  timestamps: the anchored comparisons (resolve_locktime_against_tx and
  resolve_date_to_check with built_locktime) already prevent the daily
  invalidate prompt. Absolute values still sync on a genuine automatic
  anticipation.
- Will.remove_stale_wallet_history drops stale wallet-LOCAL will placeholders
  before every (re)build in GUI and CLI so their coins are available again.
- check_willexecutors_and_heirs raises HeirNotFoundException outside the
  count_heirs gate: a shortened relative recipe on a signed will now triggers
  a plain rebuild ("no heirs" only when there really are none).
- is_locktime_below_threshold compares the settings on one reference frame
  (resolve_guard_threshold), no longer against the built-will anchor.

Tests: purge unit + call-site, no-heirs, guard, anticipated-rebuild
end-to-end (GUI) + CLI mirror + fixed_percent_lists_amount unit,
relative-preserve sync; conftest restores electrum.constants.net after each
test (cross-file pollution guard).

tests/test_core_checkalive.py

@@ -18,6 +18,7 @@ from bal.core.checkalive import (  # noqa: E402  (path insert above)
     CheckAliveError,
     check_alive_expired,
     resolve_date_to_check,
+    resolve_guard_threshold,
 )
 
 # ------------------------------------------------------------------ #
@@ -148,6 +149,76 @@ def test_advanced_mode_relative_locktime_without_built_tx_falls_back():
     assert abs(result - expected) < 1
 
 
+# ------------------------------------------------------------------ #
+# resolve_guard_threshold
+# ------------------------------------------------------------------ #
+
+
+def test_guard_threshold_basic_mode_returns_none():
+    fake_now = 1_800_000_000.0
+    threshold = resolve_guard_threshold(True, {"threshold": "30d"}, now=fake_now)
+    assert threshold is None
+
+
+def _guard_locktime(settings, fake_now):
+    """Reproduce the call-site locktime expression of the guard."""
+    from bal.core.plugin_base import BalTimestamp
+
+    return BalTimestamp(settings["locktime"]).to_timestamp(fake_now)
+
+
+def test_guard_threshold_absolute():
+    fake_now = 1_800_000_000.0
+    locktime = fake_now + 90 * 86400
+    threshold = locktime - 30 * 86400
+    settings = {"locktime": locktime, "threshold": threshold}
+    assert resolve_guard_threshold(False, settings, now=fake_now) == threshold
+
+
+def test_guard_threshold_relative_fresh_anchor():
+    """A relative threshold must be anchored to the FRESH locktime so the
+    guard and the settings always share one reference frame.
+
+    Regression for the false positive where a still-valid built will frozen at
+    a LONGER delivery ("2y") anchored ``date_to_check`` beyond the currently
+    stored shorter delivery ("1y"): the old guard compared the fresh "1y"
+    locktime against that anchored threshold and wrongly fired "locktime is
+    lower than threshold", even though the settings themselves are consistent
+    (locktime is 30d AFTER the threshold).
+    """
+    fake_now = 1_800_000_000.0
+    settings = {"locktime": "1y", "threshold": "30d"}
+    locktime = _guard_locktime(settings, fake_now)
+    threshold = resolve_guard_threshold(False, settings, now=fake_now)
+    assert threshold is not None
+    assert locktime > threshold  # internally consistent: no fire
+    assert threshold > fake_now
+    # The helper takes no built anchor: a frozen "2y" built will must NOT
+    # contaminate the result, although resolve_date_to_check (the expiry
+    # reference) legitimately keeps using it.
+    frozen_two_years = locktime + 365 * 86400
+    anchored = resolve_date_to_check(
+        False, settings, now=fake_now, built_locktime=frozen_two_years
+    )
+    assert anchored > threshold  # built anchor pushes date_to_check forward...
+    assert locktime < anchored  # ...which is exactly what used to fire the bug
+
+
+def test_guard_threshold_relative_locktime_absolute_threshold():
+    fake_now = 1_800_000_000.0
+    threshold = fake_now + 200 * 86400
+    settings = {"locktime": "1y", "threshold": threshold}
+    assert resolve_guard_threshold(False, settings, now=fake_now) == threshold
+    # "1y" from now is later than the stored absolute threshold: allowed.
+    locktime = _guard_locktime(settings, fake_now)
+    assert locktime > threshold
+
+
+def test_guard_threshold_missing_returns_none():
+    fake_now = 1_800_000_000.0
+    assert resolve_guard_threshold(False, {"locktime": "1y"}, now=fake_now) is None
+
+
 # ------------------------------------------------------------------ #
 # check_alive_expired
 # ------------------------------------------------------------------ #