Repository navigation
build(changelog): assemble CHANGELOG.md from per-PR fragments - #354
Merged
Merged
Conversation
DeepDiver1975
force-pushed
the
build/changelog-towncrier
branch
from
September 15, 2026 22:54
92e098a to
74920a4
Compare
DeepDiver1975
marked this pull request as draft
September 16, 2026 20:23
Contributor
Author
|
I am uncertain if this is the way this project shall take ... |
DeepDiver1975
commented
Sep 16, 2026
DeepDiver1975
left a comment
Contributor
Author
There was a problem hiding this comment.
@dcoapp recheck
DeepDiver1975
force-pushed
the
build/changelog-towncrier
branch
from
September 16, 2026 20:39
74920a4 to
d97be11
Compare
DeepDiver1975
force-pushed
the
build/changelog-towncrier
branch
from
September 16, 2026 21:48
d97be11 to
2ece192
Compare
Contributor
Author
Release path exercised end to endThis PR changes what
Two things worth noting for review:
|
Every pull request inserted its bullet at the top of the same list under `## [Unreleased]`, so the insertion anchor was byte-identical across branches. Git saw two independent inserts at one position, could not order them, and conflicted — and merging any one PR invalidated every other open PR's hunk. `git merge-tree` reproduces it: #334 and #335 conflicted on CHANGELOG.md and nothing else, while two PRs writing to different sections merged clean. Each change now carries its entry as its own file in changelog.d/, which towncrier folds into CHANGELOG.md when a release is cut. Different PRs create different files, so there are no shared lines left to conflict over. Fragments use towncrier's orphan naming (+<slug>.<type>.md) rather than the default <issue>.<type>.md: several entries cite the same ticket (#302), and issue-numbered filenames would have reintroduced the same collision as an add/add conflict. A `merge=union` driver in .gitattributes was tried first and rejected. It does resolve these merges correctly on the client, but GitHub does not run merge drivers server-side when computing pull request mergeability, so the PRs would still have shown as conflicting. The changelog-check job now runs `towncrier check` instead of requiring a CHANGELOG.md edit. Its job id and display name are unchanged, so an existing required status check keeps matching. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com>
The release runbook still told the maintainer to "move the [Unreleased] entries under a new version heading before tagging" — the one step this branch removes. Following it after this merges produces no section at all, and with the guard from #369 in place the tag is refused outright. Replace that with the towncrier build, keeping the three things it does not do for you: the summary paragraph, the link definitions at the bottom of CHANGELOG.md, and the milestone. The tag command moves below the assembly, which is the order it now has to happen in. CONTRIBUTING.md's Hacktoberfest bullet still listed "a CHANGELOG.md entry" among the unchanged rules, contradicting the section below it that says never to edit that file; and its "Cutting a release" section now duplicated the runbook, so it points at docs/releasing.md instead of maintaining the same steps twice. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com>
DeepDiver1975
force-pushed
the
build/changelog-towncrier
branch
from
September 20, 2026 20:55
2ece192 to
0d9f80d
Compare
DeepDiver1975
marked this pull request as ready for review
September 20, 2026 21:02
DeepDiver1975
added a commit
that referenced
this pull request
Sep 20, 2026
#354 replaced the hand-edited `## [Unreleased]` section with one file per change, so this branch's seven entries move out of CHANGELOG.md and into changelog.d/. The text is unchanged — it is the same prose these commits already carried, extracted rather than retyped. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com>
DeepDiver1975
added a commit
that referenced
this pull request
Sep 20, 2026
#354 replaced the hand-edited `## [Unreleased]` section with one file per change, so this branch's three entries move out of CHANGELOG.md and into changelog.d/. The text is unchanged — it is the same prose these commits already carried, extracted rather than retyped. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com>
DeepDiver1975
added a commit
that referenced
this pull request
Sep 20, 2026
…ation (#431) * test(acceptance): reach rail tabs by header instead of by index The workspace nav rail is about to become Stammdaten -- 0..n user-defined checklists plus per-module visibility -- so `tabs.SelectedIndex = 7` stops naming the tab it was written for. Worse, `(TabItem)tabs.SelectedItem` and `(Visual)tabs.SelectedContent` keep compiling once the rail is ItemsSource- driven and the items are view models, then throw at runtime, test by test. Centralise both in WorkspaceRenderHelper: SelectTab(window, header) finds the tab by its rendered header SelectedTabContent(window) the realized child of PART_SelectedContentHost RailHeaders(window) the rail's headers, in rail order Reading the first TextBlock under a container works for either header shape (a bare string, or the checklists' TextBlock-plus-status-dots panel) because TabControl hoists the selected content into PART_SelectedContentHost, leaving a TabItem's own subtree holding nothing but its header. Eight duplicated local Tabs(window) helpers collapse into the one. Header assertions that only restated the index just used to select the tab are gone; SelectTab throws -- naming the headers the rail does show -- when the tab is missing, which is the same check with a better failure. NavRailHelperTests pins the helper itself, including that SelectTab lands on the index it names: a silent mis-selection would otherwise weaken every module test at once without failing any of them. It also records today's rail order, which is what an empty Stammdaten layout will have to keep reproducing. No src/ change. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * feat(domain): an Einsatz carries 0..n Checklisten, not exactly two ChecklistKind had exactly two members, and an Einsatz had one item list per member. A brigade that wants no Checkliste, or five, or one called "Nachbereitung", had nowhere to put that -- and the two lists were named by hardcoded German literals in the PDF exporter and the tab headers, so the file itself never knew what its own lists were called. Incident now owns IReadOnlyList<ChecklistList>. Each list carries: Id the Stammdaten template's id, kept through seeding, so the nav layout can match an entry to the list actually present in this file Title stored bare ("Aufbau", not "Checkliste Aufbau") Keeping the title in the incident -- rather than resolving it from Stammdaten at display time -- is what lets a PDF export, a joined device and a file reopened months later name every list it contains, and what makes renaming or deleting a template harmless to an Einsatz already running. The ETB completion entry interpolates that title where it interpolated ChecklistKind.ToString() before. Because titles are bare, a migrated file keeps producing byte-identical journal text, which ChecklistListTests pins. Rehydrate's two positional checklist parameters collapse into one. That break is deliberate: a positional change nobody notices is how a caller ends up passing Abbau items as Aufbau. Every call site failed to compile and was carried over. ChecklistAufbau/ChecklistAbbau survive as computed properties over Checklists, alongside a two-list SeedChecklist overload and ChecklistDefaults.AsLists, so the persistence, sync, PDF and view-model layers are untouched here and move over one at a time. They are not [Obsolete]: TreatWarningsAsErrors would turn that into a build break at every site this commit is meant to leave alone. They go when their last caller does, and ChecklistKind goes with them. Behaviour is unchanged -- the app still seeds exactly Aufbau and Abbau. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * feat(persistence): an Einsatzdatei stores its own Checklisten (V23) An incident file recorded its Checklisten as one item table tagged with a `kind` column that had exactly two legal values, and nothing in the file said what either list was called -- "Aufbau" and "Abbau" lived as German literals in the PDF exporter and the tab headers. So the file could neither hold a third list nor name the two it had. V23 adds checklist_lists (id, ordinal, title) and checklist_items.list_id, and backfills from kind. Keeping the titles in the file is the point: a PDF export, a joined sync device and this file reopened months from now can all name every list without Stammdaten, which may since have renamed or deleted the template it came from. Deliberate choices, each pinned by a test: - A list row is created only where items exist, so a file whose Abbau list was never filled loses its permanently-empty ABBAU tab instead of carrying it forever. CHANGELOGged: it is user-visible. - list_id is nullable with no default. NOT NULL DEFAULT '<aufbau guid>' would silently file every future orphan under Aufbau; nullable still satisfies SchemaGuard's repairability rule, and the reader drops an unmatched row rather than guessing. - kind is kept, dormant, written 0. Dropping a column means rebuilding the table -- V3/V6/V17 are the only precedent -- which is not worth it for one integer nothing reads. incident_meta.ils_number made the same call. - The guid literals are Guid.ToString()'s "D" form, which is what the repository writes and Guid.Parse reads back. A format mismatch would orphan every item silently, so a test compares the two directly. - The backfill is skipped when checklist_items is absent. SchemaGuard exists because a file can reach this build missing a table this lineage assumes, and a migration that throws there leaves it unopenable. checklist_lists is registered in SchemaGuard's table list and its repairable columns, per AGENTS.md -- and, more importantly, added to Save's wipe list. Leaving it out is precisely how V21 came about: force_unit_edits was missing there and every save re-inserted the whole history, which no test caught. One here pins that saving three times leaves the row counts alone. SampleIncidentUpgradeTests runs the whole thing against a copy of the shipped docs/samples/uebung.fwincident -- a real V21 file written by a released build, not a fixture. It is copied first: opening it in place would stamp this branch's schema_version onto a tracked file. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * feat(stammdaten): 0..n Checkliste templates and a Navigation layout Stammdaten held exactly two Checkliste templates, in a table whose `ordinal` was the PRIMARY KEY -- so the two lists shared one running sequence, Abbau continuing where Aufbau left off, just to fit inside it. That scheme cannot express a third list, and nothing in it could name one. MasterDataSet now carries IReadOnlyList<ChecklistTemplate> (id, title, items) and IReadOnlyList<NavEntry>, the rail's order and per-entry visibility, over built-in modules and Checklisten alike. Module identity is a string key, not an enum. An enum persisted as an integer is exactly the mistake ChecklistKind made -- the reason this work exists -- because the numbers mean whatever the declaration order was when they were written. A key survives reordering, and one from a newer build is skipped rather than read as a different module. An empty Navigation means "this build's default" and is never materialized into a stored set. A saved copy would freeze today's module list, so a module added in a later release would stay invisible to everyone who never opened the Navigation category. For the same reason Navigation is excluded from IsEmpty: the editor offers Import only while the set is empty, so one saved layout would otherwise permanently suppress that first-run bootstrap -- the trap the Settings doc-comment already warns about. Three import shapes, because every Stammdaten file in the wild predates the newest: `checklists` wins; else the Aufbau/Abbau keys, mapped onto the frozen ids so an imported file's Aufbau list is still the one a layout names; else the oldest flat `checklistTemplate` array, every item optional. Either legacy side is dropped when empty, matching V22's "no items, no list". Export writes the new keys only -- emitting both shapes would leave two sources of truth in one file -- which is CHANGELOGged as a one-way step. The store gets md_checklist_lists / md_checklist_items / md_navigation and moves md_checklist_template across once, in C#, undoing the offset scheme with a per-list counter. It has no version marker, so the old table's own existence is the marker: it is dropped on success, the same idiom already used for md_call_signs/md_brigades. That move is the only destructive step here and gets its own transaction, since EnsureSchema otherwise runs untransacted. Verified the move tests actually bite by swapping the two lists in the mapping: three of them fail. docs/master-data.example.json gains a third Checkliste and a Navigation block with one module switched off, per AGENTS.md; the demo Stammdaten move to the new keys but stay on the default rail so the README screenshots are unaffected. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * feat(applogic): resolve the nav rail from layout plus the Einsatz's own lists The Navigation layout is global and live; an Einsatz carries its own copy of the Checklisten it started with. The two can therefore always disagree, and every way they can is a way an operator loses sight of a list they filled in. NavigationLayout.Resolve is where that is reconciled -- a pure function, so the rules are testable on their own, which is the point. empty layout -> the rail as it shipped checklist named, Einsatz has it -> shown (if visible) checklist named, Einsatz lacks it -> skipped checklist in Einsatz, not named -> appended, in the incident's own order module key this build knows -> shown (if visible) module key it does not -> skipped module/list the layout never names -> appended Two rules carry the weight. A hidden entry counts as "mentioned", so the append pass cannot resurrect a list the operator deliberately unticked -- otherwise the toggle would not work at all. And appending what the layout never named is what makes an archived Einsatz show every list it actually holds, including one created after it was closed or since deleted from Stammdaten. Writing the tests first settled a question the design had left open: omitting an entry is not how something is hidden -- `visible: false` is. Omission only happens in a hand-edited file or a layout written before that module existed, and appending is the safe reading of both. Three of my first expectations assumed otherwise and were wrong, not the code; they now state the rule directly instead of pinning the whole rail. The ETB's own visible flag is ignored rather than trusted: the editor disables that checkbox, but a hand-edited JSON can still say false, and the ETB is the legal record every system entry lands in. The final insert-at-0 guard is unreachable while NavModules.All contains it, and says so -- it is kept as the one place that guarantee is stated outright. No UI yet; nothing calls this. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * feat(workspace): build the nav rail from Stammdaten instead of ten literal tabs The rail was ten <TabItem> elements in the XAML, each bound to its own view- model property, with AUFBAU pinned to slot 0 and ABBAU to slot 9. Nothing about it could vary per brigade. It is now ItemsSource-driven over NavItems, which IncidentWorkspaceViewModel builds by handing the Stammdaten layout and the Einsatz's own Checklisten to NavigationLayout.Resolve. One header template serves every tab. The dots are the reason WorkspaceNavItemViewModel exists rather than binding the checklist through directly: IsComplete/IsIncomplete are plain booleans that are both false for a module. Binding a nullable checklist's AllMandatoryDone would produce no value on a module tab, leaving IsVisible at its default true and putting a status dot on all ten. Two things this turned up that the plan had not: - The custom TabItem ControlTemplate binds the header presenter's Content but never its ContentTemplate. Harmless while every header was a literal string or panel; with a templated header the app-level ViewLocator claimed it instead and every tab rendered "Nicht gefunden: WorkspaceNavItemView". One TemplateBinding fixes it, and the #146 ScrollViewer work is untouched. - The content presenter reuses its realized Border across tab switches and re-binds it, so visual identity is not evidence of anything. The helper test now asserts the subtree is the other module's, which is what it meant. ChecklistViewModel takes a ChecklistList instead of a ChecklistKind and resolves its items by list id -- a captured list object would go stale, since a remote broadcast replaces the whole Incident. That was the enum's last use, so ChecklistKind is deleted. ChecklistView's heading binds the list's name instead of reading "CHECKLISTE", through the same casing helper the tab uses so the two cannot disagree. Rail labels live beside the module keys as a separate map: the keys are persisted Stammdaten, these are German display strings, and they should not be able to drift into one another. The x:Named status dots (AufbauCompleteDot and friends) could not survive templating; WorkspaceRenderHelper.RailStatusDots scopes the lookup to a named tab's container instead. Screenshots regenerated: the rail is pixel-identical, only the Checkliste heading changed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * feat(stammdaten): create, rename, reorder and delete Checklisten in the editor The editor had two fixed Checkliste sections, "Checkliste Aufbau" and "Checkliste Abbau", and a comparator whose only job was to keep Aufbau above Abbau. This replaces them with as many sections as the brigade has lists, and adds the Navigation category that orders the Einsatz sidebar and switches its entries on or off. A Checkliste's section Title *is* its name, bound two-way from the detail pane, so typing renames the rail entry and its Navigation row as you go. Its Id never changes -- that is what an Einsatz seeded from it carries, and what keeps the two matched across a rename. Two rules earn their tests: - A new Checkliste is appended to the Navigation layout as well. The resolver's append rule rescues lists an *Einsatz* holds, never rows a *layout* is missing, so a list that never reached the layout would be invisible everywhere. - A blank name falls back to "Checkliste" rather than dropping the list. The operator cleared the name, not the items. Deleting asks first and says what it does not do: Einsätze already started keep their own copy, and the workspace appends any list a file holds that the layout no longer names. SectionTitleComparer is deleted outright. Checkliste order is the operator's own now, expressed in the Navigation list, and no comparer can express that -- so the Checklisten sit after the alphabetical categories, in Stammdaten order. This also retires the last transitional shims, which is most of the diff: Incident.ChecklistAufbau/Abbau, the two-list SeedChecklist, MasterDataSet's ChecklistTemplate* and ChecklistDefaults.AsLists are gone, and the sync snapshot and PDF export moved to n lists with them. A joined device now gets the host's lists by name; a PDF titles each list from the incident file rather than from a template that may since have been renamed. LocalIncidentSession.StartNew keeps only the ChecklistSeed form. The two-list convenience moved into the test projects as TestSession/TestChecklists instead: production code must say which lists it means, but a test about the CO-Messprotokoll should not have to, and ~40 such tests are untouched by it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * docs: refresh the sample Einsatzdatei and the Checklisten claim `make samples` regenerates docs/samples/uebung.fwincident, now written at schema V22 with its two Checklisten carried in the file's own checklist_lists table rather than inferred from a kind column. The README's feature table said "Aufbau- und Abbau-Checklisten", which is no longer what the app does. NOTE: that row lives in the "Warum Lagebuch?" section, which AGENTS.md requires to be kept in step with the website's src/content/docs/index.mdx in the same pull-request pair. That companion change is NOT in this branch -- it belongs to CodeForFire/codeforfire.github.io and has to be opened there before this lands, or the site will keep claiming exactly two Checklisten. Nothing the site copies was renamed: docs/datenschutz-und-sicherheit.md, docs/logo/, docs/demo/einsatz-flow.gif and docs/screenshots/*.png all keep their paths. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * test(persistence): correct the sample-upgrade test's own description `make samples` rewrote docs/samples/uebung.fwincident at V22, so the comment claiming this exercises a migration against a released build's file is no longer true of the file in the tree. Say what it now checks -- that the shipped sample opens and keeps its Checklisten -- and where the V21 path is covered. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * test(acceptance): render the configured rail end to end The feature's own proof, rendered rather than asserted on view models: three Checklisten with one placed second, three modules switched off, and the rail comes out as exactly those eight tabs. Plus the two cases the old rail could not express at all -- a Checkliste opening under its own name instead of a fixed "CHECKLISTE" heading, and a brigade that uses no Checklisten getting a rail without any. Also pins that a module tab shows neither status dot. That is the failure the nav item's two plain booleans exist to prevent: a nullable binding producing no value would leave IsVisible at its default and put a dot on every tab. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * build: fail the build on an obsolete API, and stop using TextBox.Watermark The Checkliste name field I added last week used `TextBox.Watermark`, which Avalonia has replaced with `PlaceholderText`. The build printed AVLN5001 and carried on. `TreatWarningsAsErrors` is set, so this should not have been possible -- but it is Roslyn-only. C# obsolete warnings (CS0612/0618/0619) already fail; Avalonia's XAML compiler logs through MSBuild and went straight past the gate. Worse, a XAML warning only surfaces on a *full* rebuild, so an incremental build never showed it at all. `MSBuildWarningsAsErrors` escalates AVLN5001, which is the half of the rule that actually holds. Verified in the order that proves it rather than assumes it: added the property first, with Watermark still in place, and confirmed the build turned "Avalonia warning AVLN5001" into "Avalonia error AVLN5001" and failed. Then fixed the occurrence; a full rebuild is now clean at 0 warnings. Scoped to AVLN5001 on purpose -- the AVLN class also covers warnings that are not deprecations, and escalating those is a separate decision. Theme/Styles.axaml's `TextBox /template/ TextBlock#PART_Watermark` selector is deliberately untouched: that is Avalonia's template part name, not the obsolete property, and renaming it would break the style. AGENTS.md gets the rule beside Path.Join, CA1031 and the dead-locals rule, with the full-rebuild caveat spelled out, since that is what let this one through. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * feat(stammdaten): separate the Checklisten from the fixed categories The rail listed eight fixed app categories and 0..n Checklisten the brigade wrote itself as identical rows. Nothing said which entries a Kommandant may rename and delete and which are simply part of Lagebuch; the Checklisten only happened to sort last. The rail now shows the app's categories, then a CHECKLISTEN heading with a hairline rule, then the brigade's own lists. Only the authored group is labelled -- the screen title already says what the first group is, and a second heading would repeat it. The separation is in the view model, not painted on: `Sections` keeps the eight fixed categories and a new `Checklists` collection holds the rest. One collection for both kinds is precisely why the view could not tell them apart. Three things this turned up that a XAML-only change would have shipped broken: - **Two ListBoxes over one SelectedItem cancel each other out.** SelectedItem is a two-way direct property and Avalonia writes a target change straight back with no re-entrancy guard, so the list that does not hold the selection writes null as it clears its highlight -- blanking both rails. SelectedSection now refuses null, which is honest: the editor always has exactly one section selected. Verified by removing the guard and watching the acceptance test fail; a view-model-only test passes either way, so that test drives the real controls. - **Deleting the last Checkliste crashed.** The selection was clamped into the collection it had just been removed from, so `Checklists[0]` on an empty list threw. It now picks the nearest survivor, else the entry above the group. - **Restoring the selection by index stopped meaning anything.** A Checkliste's index in its own group says nothing about its old position in `Sections`, so saving while editing "Aufbau ELW" dumped you on Einstellungen. Restore is by identity now, which also survives an import reordering. The fixed list gets a non-virtualizing panel: an Auto row measures with infinite height, but a VirtualizingStackPanel sizes from the effective viewport, which is empty on the first pass -- it would report ~25px per row and under-realize rows in headless render tests. Eight rows never need it. The add button stays docked rather than moving into the Grid. On a short rail the star row collapses first, and a clipped list is recoverable where a clipped "add" action is not. It still sits under the group, which is the last thing in the rail. Rejected: a second DataTemplate on the one ListBox, badging the Checkliste rows. No selection hazard and no test churn, but it cannot carry a heading -- and the heading is the thing that actually names the distinction. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * fix(stammdaten): give the rail a single scroll region The rail I split last commit gave the fixed categories an Auto row and the Checklisten the star row. In the running app the eight categories eat nearly the whole rail, so the Checklisten collapsed to ~90px: their own scrollbar for a single entry, and a row clipped mid-card under the CHECKLISTEN rule. Two scroll regions in a 220px column, and the group that grows without bound got the leftovers. One ScrollViewer owns the rail now. Both lists size to their content and RailScroll does all the scrolling, so there is never a scrollbar inside a scrollbar and never a half-row. What makes each list report its full height is the infinite constraint from RailScroll plus the non-virtualizing panel -- not, as I first assumed, the Disabled scrollbar setting. ScrollContentPresenter measures its child with infinity only when CanVerticallyScroll is *true*, which Disabled turns off. What Disabled buys is that the inner presenter can never build a scrollable extent, which is what keeps the wheel, touch and keyboard working: its wheel handler early-outs without marking the event handled, and BringDescendantIntoView returns false for an offset that cannot move, so the focused row scrolls into view out at the rail. Hidden would have been a trap -- it leaves the list able to scroll, silently, with no scrollbar to show where the content went. The Grid became a StackPanel: with every row Auto the Grid *was* a StackPanel, and dropping it also removes the Grid.Row overlap of the empty-state text and the list, which only worked because one is always collapsed. The rail's item template no longer repeats the horizontal padding the ListBoxItem already applies -- it was the only list in the app whose text started 26px in instead of 20. The vertical 4 stays on purpose: it holds the row's hit target at 51px, above the 48dp touch-target minimum the Android head needs. This is a cleanup, not the fix; 64px across eight rows does not close a 90px shortfall. The add button keeps its docked position and gains a hairline above it, so the slack under the last Checkliste reads as a footer. Declared after the button, because a DockPanel gives the first Dock="Bottom" child the bottom-most strip -- declaring the rule first drew it under the button, which is how I first wrote it. The new test earns its keep: my first version asserted Extent <= Viewport on each inner list, which passes whether or not the layout is right, since Disabled makes a squeezed list clip rather than scroll. Verified by constraining ChecklistList to 60px -- it passed. It now asserts each list is as tall as its own rows, which fails that mutation naming the real symptom. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * test(applogic): point tests main added at the test-only session helper Rebasing onto main surfaced two conflicts git could not see: main's #302 test for a faulting Android file pick and #399's ScbaSafetyTruppTests both call LocalIncidentSession.StartNew with the two-list Aufbau/Abbau signature, which this branch removed in favour of ChecklistSeed. Neither commit touched the same lines as this branch, so the rebase applied cleanly and the build failed instead. Same treatment as the ~40 other tests that only want an incident and are not about Checklisten: TestSession.StartNew. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * test(stammdaten): drop the redundant EditorSection cast `Sections` is an `ObservableCollection<EditorSection>`, so `FirstOrDefault` already yields `EditorSection?`; `ChecklistTemplateSection` converts to `EditorSection` implicitly, so the `??` picks the same best common type with or without the cast. CodeQL's `cs/useless-cast-to-self` is right that it buys nothing. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * docs(changelog): carry the Checklisten entries as changelog.d fragments #354 replaced the hand-edited `## [Unreleased]` section with one file per change, so this branch's seven entries move out of CHANGELOG.md and into changelog.d/. The text is unchanged — it is the same prose these commits already carried, extracted rather than retyped. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> --------- Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
DeepDiver1975
added a commit
that referenced
this pull request
Sep 20, 2026
…ment #354 replaced the hand-edited `## [Unreleased]` section with one file per change, and the rule is one entry per pull request, not one per commit — so this branch's three entries become a single `added` fragment describing the Messreihe, the ETB visibility of a changed reading and the WOHNUNG BEARBEITEN sidebar together. The prose is rewritten rather than concatenated, but nothing a reader would act on is dropped: the 600px minimum-height case where FERTIG was cut off and a Wohnung could not be saved at all, the Android variant, the six-row cap, and "files written before this open unchanged with an empty Messreihe" all survive. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com>
DeepDiver1975
added a commit
that referenced
this pull request
Sep 21, 2026
…ment #354 replaced the hand-edited `## [Unreleased]` section with one file per change, and the rule is one entry per pull request, not one per commit — so this branch's three entries become a single `added` fragment describing the Messreihe, the ETB visibility of a changed reading and the WOHNUNG BEARBEITEN sidebar together. The prose is rewritten rather than concatenated, but nothing a reader would act on is dropped: the 600px minimum-height case where FERTIG was cut off and a Wohnung could not be saved at all, the Android variant, the six-row cap, and "files written before this open unchanged with an empty Messreihe" all survive. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com>
DeepDiver1975
added a commit
that referenced
this pull request
Sep 21, 2026
…ins Protokoll bringen (#432) * feat(etb): give automatic measurements their own direction Practitioners reported that a changed CO reading "is not documented". It was: Incident.RecordCoValue has always written an ETB line for every real change. But it wrote it as EtbDirection.System, and the ETB hides system messages by default since #223 -- so the line only existed behind a checkbox nobody knew about. A CO reading is not bookkeeping. "Einsatz begonnen" is; what a Trupp measured is a professional record. Split the two: EtbDirection.Measurement (ordinal 4, appended last, because the direction is persisted by ordinal and 0-3 are a wire contract) carries the ppm line, and the "Systemmeldungen ausblenden" filter keeps testing for System alone. The reading is therefore visible by default while the trace around it stays hidden, which is what #223 actually asked for. Measurement is app-written like System, so it is kept out of the direction picker and out of reach of EditJournalEntry. Those two questions -- "may a human pick it" and "may a human rewrite it" -- now have one answer in EtbDirections.IsAppWritten instead of three copies of the same comparison. Visibility deliberately does not use it: that is the whole point of the split. AppendSystemEntry already had two optional parameters, so a trailing direction would have forced every call to read AppendSystemEntry(..., direction: ...) -- self-contradictory. The body is extracted once as AppendAutomaticEntry, with AppendSystemEntry and AppendMeasurementEntry as thin wrappers; the ~20 existing call sites are untouched. Only the ppm line moves. Adding, removing and searching a Wohnung stay System. Refs #424 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * feat(co-messung): record a measurement series per Wohnung Dwelling.CoValue was a single int? that RecordCoValue overwrote, so the previous reading was gone the moment a new one arrived. The only trace left was the ETB line -- German prose that nothing in the repo ever reads back. A Wohnung that went 120 -> 40 -> 5 ppm could not be asked what it had measured, which is why the CO section of the report shows a number and not a Verlauf. Dwelling now carries Readings: the whole run, oldest first. Deliberately the full series rather than ForceUnit's "superseded values only" pile -- a strength edit dates a correction, but a Messreihe has to date each measurement, and the reading an Einsatzleiter most wants a time for is the latest one. Keeping the current value out would leave exactly that one undated and force every renderer to special-case "and then the parent's value, at an unknown time". CoValue stays, and is now always the last reading rather than a separately tracked number. Value is nullable so clearing a reading is recorded as the correction it is: once the value is gone the Wohnung reads "kein Messwert" and the series is the only surviving record that 120 ppm was ever there. WithCoValue becomes WithCoReading, which moves the value and appends to the series as one operation. The value can no longer travel without leaving a trace; the invariant is in the type rather than in the caller's discipline. The no-op guard in RecordCoValue stays exactly where it was. It is what stops a retried or duplicated RecordCoValueCommand from inflating the series, and it is already the guard that suppresses a duplicate ETB line. Consequence recorded in a comment: a genuine re-measurement returning the identical value is not captured. Readings is appended to Rehydrate with a default, the convention ForceUnit already uses, so a file or snapshot written before this rehydrates with an empty series -- no history is invented for old incidents. Refs #424 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * feat(persistence): store the CO measurement series Schema v22 adds co_readings: one row per reading, keyed by Wohnung and ordered by an ordinal, exactly the shape force_unit_edits and scba_pressure_readings use. Deliberately one row per reading rather than one column per gas. #410 turns the single CO value into up to six named slots, and it can extend this table with gas_slot/unit through AddColumnIfMissing without touching the row shape or the series it already holds -- so the history is designed once, here, and not a second time. For the same reason there is no UNIQUE (dwelling_id, measured_at): those six slots will share one timestamp. The bump also does ApplyV7's job. A file written by this build may carry the new EtbDirection.Measurement (ordinal 4) in etb_entries.direction, and the version > CurrentVersion guard is what stops an older build rendering that unknown ordinal as "4". Nothing is backfilled. The readings a pre-v22 incident took survive only as German ETB prose, and matching that text back to a (building, floor, apartment) triple would have to guess past renamed buildings and the Links/Mitte/Rechts labels. A fabricated Messreihe in a document a Kommandant relies on is worse than an honest gap, so an upgraded file keeps its co_value and gains an empty series -- pinned by a test. co_readings is in Save's DELETE list and in SchemaGuard. The first is what ApplyV21 had to clean up after for force_unit_edits: a child table missing from that list means every save stacks another copy of the history onto the file, so Save_Twice_DoesNotDuplicateTheSeries guards it directly. The second is the anti-drift guard that keeps the column repairable in a file that skipped the migration. The sample incident now measures 2. OG links three times (120 -> 40 -> 5 ppm after ventilation) so docs/samples/uebung.fwincident demonstrates an actual Verlauf instead of a single number, and is regenerated with `make samples`. Refs #424 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * feat(sync): carry the CO measurement series in the snapshot A joined device has to see the Verlauf, not just the current number: the series is what the CO tab and the report render, so a device that joined mid-Einsatz would otherwise show every Wohnung as having been measured exactly once. DwellingDto gains Readings with a trailing default, the convention BuildingDto already uses for ApartmentLabels and ApartmentCounts. That default is this layer's whole compatibility story -- the wire carries no schema version of its own -- so a payload written before #424 has no readings key, deserializes to null and maps to an empty series rather than a null that throws on first render. Pinned by a test that first asserts the key is really there, so it cannot start passing vacuously if the property is ever renamed. RecordCoValueCommand is deliberately unchanged. It states an intent ("set this Wohnung to X"), not a state transfer: the host applies it through Incident.RecordCoValue, which mints the reading from its own clock and the command's operator, and the series then reaches every client through the ordinary snapshot broadcast. Giving the command a timestamp would let a client forge a measurement time, which is strictly worse. Attribution still follows the originating crew, and a resent command still appends nothing, because the no-op guard in RecordCoValue sees the unchanged value. Refs #424 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * feat(co-messung): show the Messreihe in the sidebar and the PDF The series existed in the model but nowhere a crew could see it. Two places now render it, and one deliberately does not. Editor sidebar: a MESSREIHE list under the ppm field, newest first, each row carrying time, value and the crew that took it. Values keep their own severity colouring through the existing coValue styles, so a reading that has since fallen still reads as having been dangerous when it was taken. The list is read-only and is not part of the edit buffer: a keystroke is not a measurement, ABBRECHEN has to be able to discard it, and the timestamp of a pending value does not exist yet -- it is minted by the clock when FERTIG commits. There is structurally nothing for ApplyPendingEditToMatrix to preview. Tile tooltip: the same run on one line, from two readings up. The 62px tile itself is untouched -- it is already tight, and #410 wants to put up to six values in it. PDF: the Verlauf indented under its Wohnung, "Messreihe: 08:14 250 ppm · 08:41 40 ppm · 09:02 5 ppm". From two readings up for the same reason as the tooltip: with one it merely repeats the ppm on the line above. Grey rather than severity-coloured, unlike the sidebar -- the report is scanned as a whole for red, and a red historical value would announce a danger that has already passed, while the current value one line up carries the coloured statement. The "Betroffene Wohnungen" roll-up keeps showing only the current value: it is a handover list, and a series per line would make the page unreadable. Both renderers format a reading through CoMeasurementLabels.ReadingLabel, so the app and the report cannot word the same series differently -- that is what the class is for. Verified on a generated PDF, since the Documents suite extracts no text: a three-reading Wohnung prints its run, a cleared one prints "09:02 gelöscht" and keeps the 120 ppm that preceded it, and a single-reading Wohnung prints no series line at all. Closes #424 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * docs(screenshots): refresh pdf-export.png for the visible Messung lines The only demo capture affected: etb.png is taken before the demo records any CO value and co-messung.png is taken with the sidebar closed, so both are byte-identical. This one is shot on the ETB tab after the readings, and now shows what #424 fixes -- three Messung rows visible with "Systemmeldungen ausblenden" still ticked. Refs #424 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * fix(co-messung): let the Wohnung sidebar scroll and anchor FERTIG Reported from practice with a screenshot: in WOHNUNG BEARBEITEN, ABBRECHEN and FERTIG are not merely awkward to reach, they are gone. Since #242 FERTIG is the single commit point of that sidebar -- every field edits a buffer and only ConfirmEditor writes status, ppm, Bewohner and Bezeichnung into the incident. A clipped FERTIG therefore means the Wohnung cannot be edited at all: a crew measures 10 ppm and has no way to save it. The sidebar is a Dock="Right" Border around a StackPanel, and a StackPanel measures against infinity, so nothing ever shrinks and the Border simply clips what does not fit. There was no ScrollViewer anywhere in it; the only one in the file belongs to the matrix, which is why the left half of the reported screenshot scrolls and the right half does not. This predates the Messreihe, and the new test says so in numbers: at the app's own MinHeight of 600, FERTIG is clipped by 15px with nothing measured at all. The Messreihe made it much worse -- 61px after one reading, 469px after 25 -- because the list had no ceiling and sat directly above the button that writes it. Every committed reading buried the next FERTIG a little deeper. Android, where MinHeight does not apply and the app ships as a companion, was worse still. The sidebar's StackPanel becomes a DockPanel: the title docks Top, the action buttons dock Bottom, and a ScrollViewer takes what is left -- the idiom MasterDataEditorView already uses six times, and what AboutView did for exactly this reason ("scroll rather than clip SCHLIESSEN off the bottom"). Anchoring the buttons rather than letting them scroll is deliberate: a commit button you have to go looking for is the same bug in a weaker form. AllowAutoHide="False" matches the matrix in this file, because an overlay scrollbar reserves no layout width and would sit on the input fields of a 320px column. The Messreihe gets its own cap of about six rows. Newest first, so what stays visible is what is current. This is a deliberate exception to the one-scrollbar-per-region rule HomeView documents -- there a nested list scrolling at all is the bug, here the bounded inner scroll is the point -- and the comment says so rather than leaving a future reader to find a violation. ABBRECHEN and FERTIG had no x:Name, which is precisely why no test ever missed them; they have names now, since AGENTS.md rules out finding controls by layout value. The new tests render at 600px, which no existing CO test did (they use 1200x700 and 1920x1032) -- that gap is why this shipped unnoticed. Refs #424 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * docs(changelog): carry the CO-Messung entries as one changelog.d fragment #354 replaced the hand-edited `## [Unreleased]` section with one file per change, and the rule is one entry per pull request, not one per commit — so this branch's three entries become a single `added` fragment describing the Messreihe, the ETB visibility of a changed reading and the WOHNUNG BEARBEITEN sidebar together. The prose is rewritten rather than concatenated, but nothing a reader would act on is dropped: the 600px minimum-height case where FERTIG was cut off and a Wohnung could not be saved at all, the Android variant, the six-row cap, and "files written before this open unchanged with an empty Messreihe" all survive. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * test(persistence): name V24 in the forward-compat comment The CO measurement migration moved from V23 to V24 in the rebase — #431 merged first and took V23 for checklist_lists — but this comment still said V23 adds co_readings. Everything else about the test is unchanged. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> * test(acceptance): start the CO sessions through TestSession #431 changed LocalIncidentSession.StartNew to take ChecklistSeed, and added TestSession.StartNew for tests that need an incident but are not about Checklisten. The two CO-Messprotokoll tests added on this branch still called the session directly with the old pair of (string, bool) arrays, so they stopped compiling once the rebase put them on top of #431. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> --------- Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This was referenced Sep 21, 2026
DeepDiver1975
added a commit
that referenced
this pull request
Sep 23, 2026
…ry (#453) The language of an entry was predicted by the date its fragment was committed, not by who reads it: 32 English entries in [0.6.0], 9 German, 2 ambiguous, switching on 2026-09-21. release.yml already said in a comment that the changelog is English and the German release body is what ELW crews read, so the drift contradicted a decision the repository had already written down. The rule now says English and one sentence. Short matters as much as English does: the sediment formed because 300 words of careful prose at 23:00 after an exercise is not a thing that happens twice, and the cheap path won. The pull request holds the detail; the entry says what changed. Every bullet opens with a link to that pull request. Nobody can write it by hand — fragments are orphans (`+<slug>.<type>.md`, so two PRs citing one ticket cannot collide), so towncrier has no number, and the number does not exist while the PR is open. After a squash-merge exactly one commit adds a given fragment and its subject ends in `(#NNN)`, so scripts/changelog-prlinks.py recovers it just before `towncrier build`. It is idempotent, and it refuses to resolve a fragment it cannot place rather than cutting a release with an entry silently unlinked. The language check is advisory on purpose. Every correct entry here carries German Fachbegriffe — Atemschutz, Stammdaten, Messreihe — so a detector strict enough to gate on would fail hardest on the entries that are right. It matches German function words only, outside code spans and links, and needs two distinct hits: against the 43 fragments of the 0.6.0 cycle that flags exactly the 9 German ones and nothing else. [0.6.0] is relinked, translated and compressed. Its links are not mechanical: 25 of its 43 fragments were backfilled by #354 and #434, so git names the changelog-infrastructure PR rather than the change, and each one was matched by hand instead. [0.1.0] through [0.5.0] stay as published — they predate changelog.d/ entirely and are a record of what shipped. Refs #447 Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The problem
Every pull request inserted its bullet at the top of the same list under
## [Unreleased]inCHANGELOG.md. The insertion anchor was byte-identical acrossbranches (
### Fixedfollowed by the same context line), so git saw two independentinserts at one position, could not order them, and conflicted — even though nothing
semantically conflicted. Merging any one PR then invalidated every other open PR's hunk,
which is the rebase party seen on #331.
Reproduced mechanically with
git merge-treeagainst today'smain— two branches eachadding one
### Fixedbullet citing the same ticket:CHANGELOG.md(today)CHANGELOG.mdwas the only conflicting filechangelog.d/(this PR)Positional, not semantic. #431 and #432 are both open and both edit
CHANGELOG.md.The fix
Each change now carries its entry as its own file in
changelog.d/, whichtowncrier folds into
CHANGELOG.mdwhen a release iscut. Different PRs create different files, so there are no shared lines left to conflict
over — and unlike a client-side merge driver, GitHub sees no conflict either.
towncrier.toml— Keep a Changelog section order,wrap = falseso hand-writtenprose is never reflowed, and
title_formatmatching the existing## [x.y.z] - dateheadings.
changelog.d/— the 42 entries currently under## [Unreleased], migratedverbatim (extracted programmatically, not retyped).
CHANGELOG.md—## [Unreleased]becomes a pointer tochangelog.d/plus thetowncrier insertion marker. Released sections and the link definitions are untouched.
changelog-checkCI job — now runstowncrier checkinstead of demanding aCHANGELOG.mdedit. Job id and display name are unchanged, so an existing requiredstatus check keeps matching. Skippable with a
no-changeloglabel.CONTRIBUTING.md,AGENTS.md, the PR template, anddocs/releasing.md(second commit), whose runbook still told the maintainer to move the
[Unreleased]entries under a new version heading by hand — the one step this PR removes.
Two decisions worth flagging
Fragments use towncrier's orphan naming
+<slug>.<type>.md, not the default<issue>.<type>.md. Several entries cite the same ticket(#302), so issue-numberedfilenames would have reintroduced the identical bug as an add/add conflict. The
+<slug>form is unique per change.
A
merge=uniondriver in.gitattributeswas tried first and rejected. It does mergethese PRs correctly and completely on the client, and rebase honours it — but GitHub does
not run merge drivers server-side when computing PR mergeability
(open, unimplemented feature request).
PRs would still have shown as conflicting. It would have downgraded the party, not ended it.
Verification
towncrier build --draftrenders all 42 entries;compared set-wise against
main's## [Unreleased]per section(7 Added / 7 Changed / 27 Fixed / 1 Security): identical content, no entry missing
or extra, umlauts, backticks, bold and inline
(#nnn)references intact.main.towncrier check— on a branch touchingsrc/with no fragment it exits 1; with oneit exits 0.
towncrier build --version 0.6.0— writes## [0.6.0] - 2026-09-20below themarker, empties
changelog.d/, leaves the## [Unreleased]pointer in place, and leavesevery released section and link definition byte-identical to
main.grep "^## \[${VERSION}\]"matches theheading towncrier's
title_formatproduces, checked on towncrier's own output ratherthan a hand-written heading.
and real GitHub Releases in a private throwaway repo — see the comment below.
Merge order
CHANGELOG.mdon their own, whichmakes "tagged without running
towncrier build" possible for the first time. ci(release): link the changelog from release notes instead of inlining it #369'sversionjob refuses such a tag in seconds, before any build minutes. Without it anotes-less release is caught only by someone noticing.
CHANGELOG.mdedits. Merge them before this, or converttheir entries to fragments — otherwise their entries are dropped in the rebase.
Notes
### Sectionheading, which the olderhand-written sections don't have. Both render identically; not worth owning a custom
template over, and older sections were left as they are.
mainat a3703f1; the only conflict wasCHANGELOG.md, which is the bugthis PR exists to remove — a last, fitting demonstration of it.
🤖 Generated with Claude Code