Skip to content

feat(atemschutz): model the Sicherheitstrupp instead of hoping for it - #408

Merged
DeepDiver1975 merged 3 commits into
mainfrom
feat/sicherheitstrupp
Sep 20, 2026
Merged

DeepDiver1975 merged 3 commits into
mainfrom
feat/sicherheitstrupp

Conversation

@DeepDiver1975

@DeepDiver1975 DeepDiver1975 commented Sep 18, 2026 •

Copy link
Copy Markdown
Contributor

Closes #399.

FwDV 7 expects a Sicherheitstrupp to be posted before a Trupp goes under air, and the
Einsatzbericht is read afterwards to see that it was. Until now grep -ri sicherheitstrupp src/
returned nothing — the concept existed only as a free-text Trupp-Art in the demo Stammdaten.

What this adds

AtemschutzTrupp gains a nullable SafetyTruppId. The link is by id, never by the German
designation
— #398 is removing the last behaviour that keys off a Trupp-Art name, and this must
not add another. A Trupp called "Sicherheitstrupp" is a naming convention; the link is the fact.

Under each Trupp the grid carries a picker of the Trupps currently Bereitgestellt. The
assignment may be changed at any time, including while the covered Trupp is under air — a
Sicherheitstrupp really is relieved mid-Einsatz, and a record that can never be corrected is worse
than one that moves.

Nothing is refused. One crew may stand by for several Trupps, because whether that is
defensible is the Einsatzleiter's call and not the file's. The app says so instead, in the same
amber non-blocking shape as RolesView's UnknownRoleHint:

Situation Hint
Under air with nobody behind it kein Sicherheitstrupp
The assigned crew went under air itself nicht mehr bereitgestellt
One crew shared across Trupps deckt auch Trupp 2

The two things that are refused are the two that can only ever be bugs: a Trupp covering itself,
and a link to a Trupp outside this incident.

The ETB carries the chronology

No second history table. Every change is a system entry, and the "im Einsatz" line now names the
Sicherheitstrupp — or states its absence outright, which is the line the Einsatzbericht is read
for:

Trupp 1 (Angriffstrupp) im Einsatz, Sicherheitstrupp: Trupp 3 (Sicherheitstrupp)
Trupp 1 (Angriffstrupp) im Einsatz, ohne Sicherheitstrupp
Sicherheitstrupp für Trupp 1 (Angriffstrupp) festgelegt: Trupp 3 (Sicherheitstrupp)
Sicherheitstrupp für Trupp 1 (Angriffstrupp) gewechselt: bisher Trupp 3 (…), jetzt Trupp 4 (…)
Sicherheitstrupp für Trupp 1 (Angriffstrupp) aufgehoben: bisher Trupp 3 (Sicherheitstrupp)
Trupp 3 (Sicherheitstrupp) geht selbst unter Atemschutz — Trupp 1 (Angriffstrupp) ist ohne Sicherheitstrupp

That last hook is needed only at Start: standing by means bereitgestellt, and Start is the
sole way out of that state (Withdraw requires IsActive, MarkRemoved requires
IsWithdrawing), so the other two transitions would only repeat a loss already recorded.

The picker is a flyout, and that is the interesting part of this PR

It started as a ComboBox with SelectedItem bound two-way. Running the app killed it:

System.ArgumentOutOfRangeException: Index was out of range
  at Avalonia.Controls.ItemsSourceView`1.GetAt(Int32 index)
  at …Selection.SelectedItems`1.GetEnumerator()+MoveNext()
  at …SelectionModel`1.EndBatchUpdate()
  at …SelectingItemsControl.UpdateSelection(…)   ← nested
  at …SelectingItemsControl.UpdateSelection(…)   ← outer, from the click
  at Avalonia.Controls.ComboBox.UpdateSelectionFromEvent

The two nested UpdateSelection frames are the whole story. Avalonia commits a selection inside a
BatchUpdateOperation holding indices into the ItemsSource. The setter wrote to the
incident, that write came back through Changed, and the refresh both re-wrote SelectedItem and
mutated the very collection being indexed.

Three attempts, each moving the symptom rather than removing it:

Attempt Result
Trupps.Clear() + re-add (the original) ComboBox destroyed mid-selection, popup stranded holding the pointer grab — frozen app
Reconcile rows in place (b30ccec) Exposed that each update handed over a new options list, so Avalonia reset the selection and wrote — kein — back — silent data loss, the assignment wiped moments after it was made
Stable ObservableCollection mutated in place (b30ccec) Removed the data loss, made the crash deterministic

So the selection model is gone rather than timed around: a two-way-bound SelectedItem cannot
drive a mutation that feeds back into its own ItemsSource.
The picker is now a button and a
Flyout whose entries are commands — the idiom ForcesView already uses for the Stärke editor. A
command runs after the flyout dismisses, with no SelectionModel, no ItemsSourceView and no
write-back.

Deferring instead was checked and would not have worked: IUiDispatcher exists and AppLogic can
reach it, but AvaloniaUiDispatcher.Post deliberately runs inline when already on the UI
thread, so posting from a setter would not have unwound the event.

Row reconciliation from the second attempt is kept — it is what #294 asks for, and it stops a row
being torn down while the operator is using it.

Screenshots

The picker stacks under the Trupp name rather than taking a column of its own. Eight columns
already need every pixel of a 1280px ELW screen — LayoutAlignmentTests pins that — and the width
for a ninth had to come from somewhere. The first attempt took 27px from AKTION and clipped the
RÜCKZUG button
, so that version was abandoned. Vertical space is the one thing this grid has
spare. TRUPP grows 170 → 190px; no other column moves, and the Trupp name now fits instead of
truncating.

Vorher

Nachher

The PDF makes the same trade

A4 portrait less the 1.5 cm margins leaves ~510pt, of which the five constant columns already
claim 345 — so Mannschaft is ~71pt and already wraps. A ninth column would cut it to 55pt
(relative) or 45pt (constant). The Sicherheitstrupp therefore rides as a grey sub-line in the
Trupp cell, costing no width:

Trupp 1 (Angriffstrupp)
Si.-Trupp: Trupp 2          ← 8pt grey

(#404 — the PDF is A4-only with hardcoded column widths — is what would make a real column
affordable later.)

Storage and sync

Schema V22 adds a nullable safety_trupp_id; additive, so an older file reads back with no
Sicherheitstrupp recorded, which is exactly what was true of it. SchemaGuard's expected-column
table (from #396) learns it too.

A stored id is deliberately never validated — the same stance Rehydrate already takes on a crew
it cannot validate. A dangling or unparseable value degrades to "none" rather than refusing to
open the Einsatz. A live command carrying an unknown id is a different thing and is refused.

Joining already demands an exact app-version match, so the new setScbaSafetyTrupp command needed
no wire shim.

Tests

1437 pass, 48 new. Domain (13), ViewModel incl. all six ETB wordings and the re-entrancy guards
(21), sync serialization + applier + snapshot (5), persistence round-trip, a V21 file upgrading, a
corrupt id and a dangling id (4), plus the acceptance renders.

Three of those drive the real control end to end — open the flyout, invoke an entry, assert the
incident took it, confirm Escape dismisses it. The ComboBox design could not be tested this way
at all:
its failure lived inside Avalonia's selection commit, which headless cannot reproduce
even with simulated clicks. That is the concrete argument for the control change, beyond the crash
itself.

docs/samples/uebung.fwincident and the screenshots/GIF are regenerated; the demo Angriffstrupp
now has a Sicherheitstrupp posted in the correct FwDV 7 order.

Not in this PR

CoMessprotokollView's BuildingOptions is the same hazard — an ObservableCollection cleared
and refilled under a two-way SelectedBuilding. They have already been bitten by a sibling of it
(see the comment at CoMessprotokollViewModel.cs:457-467) and worked around it. It avoids the
index crash only because Clear() raises a Reset rather than Replace/Remove. Worth its own
issue alongside #294.

🤖 Generated with Claude Code

DeepDiver1975 and others added 3 commits September 20, 2026 22:01
FwDV 7 expects a Sicherheitstrupp to be posted before a Trupp goes under
air, and the Einsatzbericht is read afterwards to see that it was. Until
now `grep -ri sicherheitstrupp src/` returned nothing: the concept lived
only as a free-text Trupp-Art in the demo Stammdaten. There was no link
from a deployed Trupp to the crew standing by for it, nothing separating
a reserve crew from one merely bereitgestellt, and nothing in the report.

`AtemschutzTrupp` gains a nullable `SafetyTruppId`. The link is by id,
never by the German designation — #398 is removing the last behaviour
that keys off a Trupp-Art name and this must not add another. Under each
Trupp the grid now carries a picker of the Trupps currently
bereitgestellt, and the assignment may be changed at any time, including
while the covered Trupp is under air: a Sicherheitstrupp really is
relieved mid-Einsatz, and a record that can never be corrected is worse
than one that moves.

Nothing is refused. One crew may stand by for several Trupps, because
whether that is defensible is the Einsatzleiter's call and not the file's.
The app says so instead: "kein Sicherheitstrupp" once a Trupp is under air
with nobody behind it, "nicht mehr bereitgestellt" when the assigned crew
goes under air itself, and "deckt auch Trupp N" when one crew is shared —
the same amber, non-blocking shape as RolesView's UnknownRoleHint. The two
things that are refused are the two that can only ever be bugs: a Trupp
covering itself, and a link to a Trupp outside this incident.

The ETB carries the chronology, so no second history table is introduced.
Every festlegen, wechseln and aufheben is a system entry, the "im Einsatz"
line names the Sicherheitstrupp or states its absence outright, and a
standby crew that goes under air itself reports the loss for each Trupp it
covered. Start is the only place that hook is needed: standing by means
bereitgestellt, and Start is the sole way out of that state.

The picker stacks under the Trupp name rather than taking a column of its
own. Eight columns already need every pixel of a 1280px ELW screen, and
the width for a ninth would have come out of the crew column or the action
buttons — the first render of that version clipped RÜCKZUG. Vertical space
is the one thing this grid has spare. The PDF makes the same trade for the
same reason: a ninth column there would cut Mannschaft from 71pt to 55pt,
so the Sicherheitstrupp rides as a grey sub-line in the Trupp cell.

Schema V22 adds a nullable `safety_trupp_id`, additive so an older file
reads back with no Sicherheitstrupp recorded — which is what was true of
it. A stored id is never validated: a file must stay openable, so a
dangling or unparseable value degrades to "none" rather than refusing the
Einsatz. Joining already demands an exact version match, so the new
command needed no wire shim.

Closes #399

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com>
… freezing the grid

The picker wrote to the domain from inside its own selection event, and
the refresh that write set off replaced both the row object and the
options list. Two things followed, and the second one is what the report
showed.

`RefreshTrupps` cleared and re-added the whole collection on every change
(#294). `Clear()` raises a Reset, so Avalonia tears down and re-realises
every container — including the ComboBox that was mid-selection, whose
popup is a real top-level window holding the pointer grab. Orphaned, it
stays on screen and keeps taking the clicks, and the app is unresponsive.
Rows are now reconciled by id and updated in place, which is what
`EtbViewModel.Sync` already does and what #294 asks for.

Worse, and silent: each update also handed the ComboBox a brand-new
options list. Avalonia reacts to a new ItemsSource by dropping its
selection and writing back — first null, then `— kein —`. The guard only
treated null as spurious, so the `— kein —` was read as the operator
choosing "no Sicherheitstrupp" and wiped the assignment a moment after it
was made. The options are now one collection for the row's life, mutated
in place; because the options are records, an entry that has not really
changed compares equal and the ComboBox sees no change at all.

The instrumented run that settled it:

    SETTER trupp=2 value=Trupp 3  domainId=      -> assign   (lands)
    SETTER trupp=2 value=(null)   domainId=712e  -> inert
    SETTER trupp=2 value=— kein — domainId=712e  -> assign   (wipes it)

Headless Avalonia cannot reproduce the freeze — it has no real popup
window to strand — but the data loss reproduces exactly, and the demo
flow caught it. Five tests cover both halves: the row and the options
collection must survive an assignment and any later change, and the
assignment must still be there afterwards.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com>
…ComboBox

Running the app and choosing a Sicherheitstrupp killed it outright:

    System.ArgumentOutOfRangeException: Index was out of range
      at Avalonia.Controls.ItemsSourceView`1.GetAt(Int32 index)
      at …Selection.SelectedItems`1.GetEnumerator()+MoveNext()
      at …InternalSelectionModel.OnSelectionChanged
      at …SelectionModel`1.EndBatchUpdate()
      at …SelectingItemsControl.UpdateSelection(…)   ← nested
      at …SelectingItemsControl.UpdateSelection(…)   ← outer, from the click
      at Avalonia.Controls.ComboBox.UpdateSelectionFromEvent

The two nested UpdateSelection frames are the whole story. Avalonia commits a
selection inside a BatchUpdateOperation holding indices into the ItemsSource.
Our SelectedItem setter wrote to the incident, that write came straight back
through Changed, and the refresh both re-wrote SelectedItem and mutated the
very collection the batch was indexing. On commit the indices no longer
matched.

This was the third failure of one design, each fix moving the symptom: clearing
the rows stranded the popup and froze the app; reconciling the rows exposed a
replaced options list that reset the selection and silently wiped the
assignment; a stable collection mutated in place removed that and made the
crash deterministic. A two-way-bound SelectedItem cannot drive a mutation that
feeds back into its own ItemsSource, so the selection model is gone rather than
timed around.

The picker is now a button and a Flyout whose entries are commands, the idiom
ForcesView already uses for the Stärke editor. A command runs after the flyout
dismisses, with no SelectionModel, no ItemsSourceView and no write-back.
SelectedSafetyTrupp is deleted outright; SafetyTruppChoice carries its own
command so the item template binds against the entry and never reaches out of
the popup's visual tree. The read-only guard moves into the ViewModel, where it
also covers a programmatic Execute — the rule ForceRow's setters already state.

Deferring instead was considered and rejected on the code: IUiDispatcher exists
and AppLogic can reach it, but AvaloniaUiDispatcher.Post deliberately runs
inline when already on the UI thread, so posting from a setter would not have
unwound the event and would have crashed identically.

Three acceptance tests now drive the real control end to end — open the flyout,
invoke an entry, assert the incident took it, and confirm Escape dismisses it.
The old design could not be tested this way at all: its failure lived inside
Avalonia's selection commit, which headless cannot reproduce even with
simulated clicks. The Rückzug-alarm row keeps its amber "kein Sicherheitstrupp"
warning, and the eight columns still fit a 1280px ELW screen.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com>
@DeepDiver1975
DeepDiver1975 merged commit 20863c2 into main Sep 20, 2026
6 checks passed
@DeepDiver1975
DeepDiver1975 deleted the feat/sicherheitstrupp branch September 20, 2026 20:14
DeepDiver1975 added a commit that referenced this pull request Sep 24, 2026
The module table, the comparison and the Probefahrt still described the
CO-Messprotokoll with a single ppm value and the sample incident with
"zwei Atemschutztrupps". Since 0.6.0 every Wohnung keeps a Messreihe whose
history goes into the ETB and the PDF (#432), and the sample incident holds an
Angriffstrupp with an assigned Sicherheitstrupp (#408) and a Wohnung measured
three times.

Same wording as on codeforfire.github.io
(CodeForFire/codeforfire.github.io#5).

Co-Authored-By: Claude Opus 5.5 (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 24, 2026
….0 (#479)

The module table, the comparison and the Probefahrt still described the
CO-Messprotokoll with a single ppm value and the sample incident with
"zwei Atemschutztrupps". Since 0.6.0 every Wohnung keeps a Messreihe whose
history goes into the ETB and the PDF (#432), and the sample incident holds an
Angriffstrupp with an assigned Sicherheitstrupp (#408) and a Wohnung measured
three times.

Same wording as on codeforfire.github.io
(CodeForFire/codeforfire.github.io#5).

Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
DeepDiver1975 added a commit that referenced this pull request Oct 1, 2026
…hem (#531)

Every session.Changed -- a save in any tab, or a change pushed from a
joined device -- cleared the Kräfte grid and re-added every row. The
Reset made Avalonia re-realise each container, so the selected row lost
its highlight and the STATUS combo lost keyboard focus right after the
status it had just written.

RowReconciler now brings a bound collection in line with the incident
by id: rows whose unit is gone are removed, new ones inserted, rows out
of place moved, and every kept row updated in place. ForceRow.Update
carries an echo guard so a value pulled from the incident is not
written back as a local edit, keeps a half-typed Stärke correction
unless the incident's Stärke itself moved, and replaces StatusOptions
only when the list really differs (the #408 trap).

Funktionen and Aufgaben follow in their own pull requests.

Refs #294

Signed-off-by: Thomas Müller <1005065+DeepDiver1975@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

feat(atemschutz): the Sicherheitstrupp is not modelled at all

1 participant