i18n review fixes (#63) and invalidation/Check fixes (#64-#67)

#63: Truman's i18n review (release gate runs i18n/CLI tests, Babel check, ruff known-first-party, no translated text stored in the wallet). #64: invalidation pays the current network fee (2-block estimate). #65: Check KeyError on a stale father link. #66: invalidation cascade with list-shaped children. #67: deselected will-executor (delete allowed if never received; broadcast skips it).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

CHANGELOG.md

@@ -3549,3 +3549,236 @@ delivery; when it is not, the plugin raises an explicit error (see below).
   catalogs refreshed (removed format strings marked obsolete).
 
 **Outcome:** DONE.
+
+---
+
+## 63. i18n review fixes (Truman's review of the translation PR)
+
+**Date:** 2026-10-07
+
+**What changed:**
+
+- `make-release.sh`
+  - Step 3 now also runs `tests/test_i18n.py`, `tests/test_translations.py`,
+    `tests/test_cli_*.py` and `tests/test_import_will_details.py` (the i18n
+    tests were never run by the release gate; the CLI and will-import tests
+    were not either).
+  - Checks up front that Babel is importable by `python3` and stops with
+    "pip install babel" otherwise: the i18n tests (step 3) and `build_zip.py`
+    (step 5) both need it, so a release no longer stops half-way.
+- `AGENTS.md`: Babel must be installed in the runtime env used for releases.
+- `pyproject.toml`: `[tool.ruff.lint.isort] known-first-party = ["bal"]`.
+  WHY: the reported I001 in `tests/test_translations.py` is not reproducible
+  from the repository root (ruff 0.15.11 and 0.6.9: clean), but it appears as
+  soon as ruff classifies `bal` as third-party (then 20 files report I001).
+  Removing the blank line would break the check where `bal` is first-party;
+  declaring it makes the result the same on every machine.
+- `bal/core/willexecutors.py` - `get_willexecutor_transactions` returns COPIES
+  of the executor records. It used to write `txs`, `txsids` and the
+  translated `broadcast_status` into the will item's own `we` dict, which is
+  saved in the wallet (stored data depended on the UI language).
+- `bal/core/will.py` - `WillItem` drops those transient fields when loading a
+  will saved by an older version, and maps a translated leading "New"
+  ("Nuovo", "Nuevo" - saved by i18n test builds) back to "New".
+- `bal/gui/qt/window.py`, `bal/cli/controller.py`: the status history of a new
+  will item is saved as the literal "New" (was `_("New")`).
+- `tests/test_core_status_language_neutral.py` (new, 7 tests; 5 fail on
+  `main`): English "New" stored, translated "New" mapped back, the mapping
+  covers every catalog's translation of "New", the push helper does not
+  modify the will item, old transient fields dropped on load.
+
+**Verification:** extended release test list (as above) plus the new test
+files: 489 passed, 1 failed (pre-existing: `test_group_e_mock_karen7::test_e5`
+needs `tests/karen7`, not in the repository; `main` 827a1d1 gives 456 passed,
+same failure). `ruff check bal/ tests/`: the same 3 pre-existing I001 as on
+`main` (`dialogs.py`, `test_core_animated_qr.py`, `test_core_qr_transfer.py`),
+none new.
+
+**Outcome:** DONE (pending the owner's test of the ZIP and Truman's review).
+
+---
+
+## 64. Invalidation pays the current network fee (2 blocks) instead of `baltx_fees`
+
+**Date:** 2026-09-26 (ported to `main` 827a1d1 on 2026-10-07)
+
+**Goal (owner request):** the will's `baltx_fees` (default 20 sat/vbyte) is
+deliberately high because the inheritance transactions may be broadcast by the
+will-executors years from now, when network fees may be much higher. The
+on-chain INVALIDATION, instead, is sent TODAY: it spends the will's coins back
+to the wallet so the old pre-signed inheritance transactions become unusable.
+It must therefore pay the current network fee, not the inheritance rate. After
+the invalidation the plugin rebuilds the new inheritance with `baltx_fees` as
+before (unchanged).
+
+**What changed:**
+
+- `bal/core/will.py`
+  - New constant `INVALIDATION_FEE_ETA_BLOCKS = 2` (owner decision: confirm
+    within 2 blocks, Electrum's own default target).
+  - New `Will.get_invalidation_fee_rate(wallet, fallback_fee_per_byte)`:
+    returns Electrum's live estimate for 2 blocks
+    (`network.fee_estimates.eta_target_to_fee`, sat/kvbyte → sat/vbyte), or
+    the fallback (the will's `baltx_fees`, i.e. the previous behaviour) when
+    there is no usable estimate: offline wallet, estimates not received yet,
+    network error or an invalid value. Never raises.
+  - `Will.invalidate_will`: uses that rate; the total fee is rounded UP to
+    whole satoshis (network rates can be fractional); logs the chosen rate and
+    its source (`network` / `fallback`). The `fees_per_byte` argument is now
+    only the fallback, so every caller (Tools → Invalidate, wizard
+    postpone/expired paths, automatic invalidation, CLI `bal_will_invalidate`)
+    gets the new behaviour without being changed. Docstrings added.
+- `tests/test_core_invalidation_fee.py` (new): rate selection (2-block target,
+  network value used as is also above `baltx_fees`, fractional rates, fallback
+  when offline / no estimate / error / invalid values) and the fee actually
+  deducted by `invalidate_will` (network vs fallback, rounding up).
+
+**Why no special case for an already-expired will:** discussed with the owner
+and intentionally left out. Before its locktime the old inheritance tx cannot
+enter a block or the mempool, so the invalidation has no competitor. RBF stays
+enabled on the invalidation, so its fee can be raised from Electrum's History
+("Increase fee") if it ever gets stuck.
+
+**Verification:**
+
+- `tests/test_core_invalidation_fee.py`: 12 passed (pytest and standalone).
+- `make-release.sh` test list + `tests/test_cli_*.py`: 435 passed, 1 failed.
+  The failure is pre-existing and unrelated: `test_group_e_mock_karen7::test_e5`
+  needs `tests/karen7`, a wallet file that is not in the repository.
+  `tests/test_import_will_details.py`: 17 passed.
+- `ruff check` clean on the touched files.
+- Live, Electrum 4.8.2 testnet, offline daemon: `bal_will_invalidate` →
+  "invalidation fee: 20 sat/vbyte (fallback), total 2200 sat" (unchanged
+  behaviour when offline).
+- Real Electrum wallet + real `FeeTimeEstimates` (1 input, 110 vbytes):
+  market 3 sat/vbyte → 330 sat; 1.5 → 165 sat; 0.4 → 110 sat (Electrum clamps
+  to its 1 sat/vbyte minimum relay fee); 45 → 4950 sat; no 2-block estimate →
+  fallback 20 sat/vbyte (2200 sat).
+- API checked in Electrum 4.7.2, 4.8.0 and 4.8.2
+  (`FeeTimeEstimates.eta_target_to_fee`, `network.fee_estimates`).
+
+**Notes:** on regtest Electrum returns a fixed 150 sat/vbyte when it has no
+estimate, so regtest invalidations will pay that rate.
+
+**Outcome:** DONE (pending the owner's test of the ZIP).
+
+---
+
+## 65. Check crashed with `KeyError: '<txid>'` (stale "father" link in the will tree)
+
+**Date:** 2026-09-27
+
+**Symptom (owner's mainnet wallet, 23 will items, Electrum 4.8.2):** the Check
+wizard stopped at "Checking your will" with `Error: 'a65a682f…'`, and the
+window refresh logged "error while updating window" with the same txid:
+
+```
+File "bal/core/will.py", line 657, in check_invalidated
+    or willtree[w.father].get_status("CONFIRMED")
+KeyError: 'a65a682f2975f0dffd9eccce9392833f227badf617fc67815674736d594d4115'
+```
+
+**Root cause:** every will item stores its parent will tx in `father`, which is
+persisted in the wallet (`"Father"` in `WillItem.to_dict`). `Will.add_willtree`
+only ever SET it (`if not father: father = ...`) and never cleared it, so when
+the parent tx was later removed from the will (rebuild, merge, replacement) the
+child kept pointing to a txid that no longer exists. `check_invalidated` then
+indexed the will with it. Reproduced on unmodified `main` (not related to #64).
+
+**What changed:**
+
+- `bal/core/will.py`
+  - `Will.add_willtree`: resets `father` on every item before rebuilding the
+    links, so the tree is always recomputed from the transactions themselves
+    (a stale value loaded from the wallet is discarded). Docstring added.
+  - `Will.check_invalidated`: defensive lookup - a father that is not (or no
+    longer) in the will is treated as "no father" instead of raising.
+- `tests/test_core_willtree_father.py` (new): stale father discarded / replaced
+  by the real parent, no crash in `check_invalidated` (with and without the
+  tree rebuild), real parent/child link built, child of an unconfirmed parent
+  still skipped. 4 of the 6 tests fail on `main` and pass with the fix; the 2
+  "normal behaviour" tests pass on both.
+
+**Verification:** make-release test list + `tests/test_cli_*.py` +
+`tests/test_import_will_details.py`: 458 passed, 1 failed (pre-existing,
+unrelated: `test_group_e_mock_karen7::test_e5` needs the missing `tests/karen7`
+wallet file). `ruff check` clean on the touched files.
+
+**Found while writing the tests, NOT fixed here (reported to the owner):**
+`Will.set_invalidate` iterates `children.items()`, but `add_willtree` builds
+`children` as a LIST, so invalidating an item that has child will txs raises
+`AttributeError: 'list' object has no attribute 'items'` (reproduced on
+`main`). The existing tests in `test_core_will_invalidate.py` set `children` to
+dicts by hand, so they do not see it.
+
+**Outcome:** DONE (pending the owner's test of the ZIP).
+
+---
+
+## 66. Invalidation cascade crashed on real will trees (`children` is a list)
+
+**Date:** 2026-09-27
+
+**Problem (found while fixing #65, reproduced on `main`):**
+`Will.set_invalidate` iterated `children.items()`, but `Will.add_willtree`
+builds `children` as a LIST of `[child_id, input_idx, prevout_idx]`, so
+invalidating a will tx that has child will txs raised
+`AttributeError: 'list' object has no attribute 'items'`. The existing tests
+set `children` to dicts by hand, so they never saw it.
+
+**What changed:** `bal/core/will.py` - `set_invalidate` accepts both shapes
+(list from `add_willtree`, dict from older data/tests) and only recurses into
+children that are still in the will. `tests/test_core_willtree_father.py`: new
+test `test_set_invalidate_cascades_to_children_built_by_add_willtree` (fails
+on `main`, passes with the fix).
+
+**Verification:** 459 passed, 1 failed (pre-existing missing `tests/karen7`);
+`ruff check` clean.
+
+**Outcome:** DONE (included in the owner's test ZIP of 2026-09-27).
+
+---
+
+## 67. Deselected will-executor: its tx can be deleted and is no longer pushed
+
+**Date:** 2026-09-27
+
+**Problem (owner, mainnet wallet):** the will had a signed tx for
+`we.divinaly.support` whose push had failed ("Push failed", "Not on server").
+After deselecting that executor, Check still reported the will as valid, the
+row stayed, and "delete" in the context menu was disabled (valid + signed
+txs were never deletable). The broadcast also still targeted the deselected
+executor, because each will item keeps its own build-time copy of the
+executor record (`selected` = True).
+
+**What changed (owner chose option "B"):**
+
+- `bal/gui/qt/lists.py` - `_can_delete`: a valid, signed tx is now deletable
+  when its will-executor NEVER received it (not PUSHED and not CHECKED). It
+  stays protected when the executor holds it (deleting it locally would not
+  cancel it on the server) and when it has no executor (backup tx).
+- `bal/core/willexecutors.py` - `get_willexecutor_transactions(...,
+  current_willexecutors=None)`: when the current executor list is given, an
+  executor that is no longer selected (or no longer listed) is skipped.
+- `bal/gui/qt/window.py` (manual broadcast) and `bal/cli/controller.py`
+  (`bal_will_broadcast`) pass the current list read from the config. The
+  wizard already filtered by the current selection.
+- Tests: `tests/test_gui_will_menu.py` (+3: deletable when never received,
+  protected when pushed/checked, protected backup tx) and new
+  `tests/test_core_deselected_executor.py` (4).
+
+**Not changed (proposal for Truman, option "A"):** Check still does not treat
+a tx of a deselected executor as "not coherent" (no automatic rebuild).
+
+**Verification:** 483 passed, 1 failed (pre-existing missing `tests/karen7`);
+`ruff check` clean on the touched files. Live (Electrum 4.8.2 testnet, CLI):
+will built for 2 executors, one deselected → `bal_will_broadcast` targets only
+the selected one (before the fix it targeted both).
+
+**Outcome:** DONE (pending the owner's test of the ZIP).
+
+**Porting note (entries 64-67):** first written on `main` 4a33116 as
+58-61 (numbers then taken by the i18n and Check Alive work), re-applied on
+`main` 827a1d1 on 2026-10-07 together with #63: extended release test list
+489 passed, 1 pre-existing failure (missing `tests/karen7`); `ruff` no new
+errors.