This is the pre-per-day session narrative from the 2026 revival, preserved verbatim from the old single-file session-log.md. These sessions were tracked by number (s13, s13b … s13z11, plus s14) rather than by calendar date — the work was never committed (HEAD was the corrupted commit until the 2026-06-06 recovery), so it has no reliable per-day dates. Newest first.

Dated work from 2026-06-06 onward lives in the per-day files (2026-06-NN.md).

The recurring theme of sessions 13f–13j: QML garbage-collects parentless C++ QObjects. Two distinct fixes came out of it — DB::qmlProtect for DB entries (13f) and qmlCppOwned() for savefile Q_INVOKABLE returns (13g–13h). See ../reference/qt-patterns.md → "QML garbage-collects parentless C++ QObjects" for the full rule.

Session 13z10–13z11 — Trainer Card layout pass (QML-only, no rebuild)

(Pre-corruption; the last pre-corruption work was s13z11.) agreed, live-iterated: (1) restored the centered grey box (the SwipeView had been changed to anchors.fill: parent in the refactor → back to centered 500×250); (2) compacted every field via one CardFront.fieldH (28) knob + tighter row margins, fixing the too-tall Qt 6 Material boxes; (3) fixed the playtime clock — narrowed each digit field to ~2 chars (PlaytimeEdit.digitPad) and stopped the row sizing to the Material implicit height (it now pins to fieldH and vcenters the :), which had made the digits ride high; (4) vcentered DefTextEdit's built-in label (it lacked verticalAlignment, so ID/Money/Coins labels rode high).

Session 13j — Pokémon box hover name finally renders (+ pen icon restored)

Hover name on the Pokémon box grid was still blank for all mons (nicknamed or not), even after the accent pill painted. Data was fine (the edit screen shows the name). Root cause: the Material Button's built-in icon+text label was not rendering its text at the button's small fixed size (height: 20, padding: 0, display: TextBesideIcon). A plain Text in the same delegate (the L## level badge) renders fine, so:

Replaced the Button's built-in label with an explicit contentItem — a centered Row of the pen icon + a Text bound to editBtn.text. Renders reliably.
Pen SVG tinted to textColorLight via MultiEffect { colorization: 1.0 } (added import QtQuick.Effects).

File: PokemonBoxView.qml. Pure QML (no rebuild).

Session 13i — Hover name attempt 1: accent pill + species fallback

First pass at the blank hover name. The label is a flat Material button; flat buttons don't paint Material.background, so there was no accent fill and (we thought) no contrast for the light text. Removed flat: true and added Material.elevation: 0 (keep the accent fill, no drop shadow). Result: the accent pill painted but the name STILL didn't show → ruled out contrast (the real fix was the contentItem in 13j). Also added a species-name fallback to getMonNickname() so un-nicknamed mons show their species (matches the in-game display). Pure QML.

Session 13h — Systemic Q_INVOKABLE-GC fix (qmlCppOwned across all …At())

Did the full systemic fix for the Q_INVOKABLE-GC bug (setObjectOwnership was chosen over parenting — parenting would have fought the existing manual deleteLater/cross-box-relocate lifecycle and risked double-frees; setObjectOwnership leaves the C++ lifecycle untouched).

Added savefile/src/pse-savefile/qmlownership.h with template<typename T> T* qmlCppOwned(T*) → QQmlEngine::setObjectOwnership(o, CppOwnership). (Qt6::Qml was already linked into savefile.)
Wrapped all 13 Q_INVOKABLE …At() returns across 12 files: connAt, grassMonsAt, waterMonsAt, signAt, spriteAt, warpAt, HoF pokemonAt, itemAt, movesAt, storage-box pokemonAt, recordAt, boxAt, partyAt. Verified none left unwrapped.
(The two app-model wrappers getBoxMon/getPartyMon were already done in 13g.)

Removes the whole class of "QML frees a savefile object I still hold" crashes/decay. Needs a rebuild. Standing rule added (CLAUDE.md + qt-patterns): wrap any new Q_INVOKABLE QObject return in qmlCppOwned().

Session 13g — Clicking-Pokémon crash root-caused (Q_INVOKABLE ownership)

Project-debugger stack trace: crash at pokemonstoragemodel.cpp:146 return !mon->isBoxMon();, read access violation at 0xffff…ffff (freed mon). Cause: getBoxMon/getPartyMon are Q_INVOKABLE and return a parentless PokemonBox; QML gives Q_INVOKABLE returns JavaScriptOwnership, so after the details editor closes QML's GC frees the mon, leaving a dangling pointer in the box's vector → next role read crashes. Fixed those two with QQmlEngine::setObjectOwnership(mon, CppOwnership); identified the systemic ~13-method version (fixed in 13h). The earlier "random terminated abnormally with no output" crashes were this same use-after-free.

Session 13f — Name-disappearing glitch = QML GC of FontDBEntry; DB::qmlProtect wired in

Repro clues: all font rendering (trainer name, full keyboard, hover tooltips) goes blank at once after "clicking around", name stops saving at the same moment, and only an app reboot fixes it. Root cause: QML was garbage-collecting the shared FontDBEntry objects (parentless, in FontsDB's vector) → dangling pointers → all font rendering AND name saving break (saving runs the name through the same font store via FontsDB::convertToCode); reboot reloads the DB. The fix machinery DB::qmlProtect(engine) existed (cascades CppOwnership to every entry in all 26 sub-DBs) but was never called — wired it into MainWindow::injectIntoQML() (+#include <pse-db/db.h>). Needs a rebuild.

Session 13e — Trainer-card randomize, TilesetPicker null, overlap, item centering, PP width

Pokémon click works (the earlier "crash" was interference). Fixes:

Randomize name on Trainer screen — NameDisplayMenu.qml (tileset version) still called the nonexistent randomName(); → randomExample(). (No-tileset version was fixed in s10.)
TilesetPicker.qml:111 Cannot read property 'name' of null — guarded the fontAt() result.
Coins/Starter/Money overlap (CardFront.qml) — fields anchored with fixed offsets from the previous field's top (40/25px) < the Qt 6 Material field height → overlap. Switched to anchor below the previous field's bottom (auto-adapts).
Item count not vertically centered — replaced fixed topPadding: 13 with verticalAlignment: TextInput.AlignVCenter.
Moves PP field too narrow (PokemonMoveSel.qml) — width now includes field padding.

Session 13d — Pokémon cell click; vertical-overlap diagnosis; é note

Pokémon click did nothing — the cell MouseArea had no onClicked (only the hover button did). Added onClicked to the cell (guards placeholder "+" slots). PokemonBoxView.qml.
Diagnosed the trainer-card/playtime vertical overlap as the Qt 6 Material control-height issue (hardcoded offsets assume the shorter Qt 5 field). Not caused by the width fixes.
é in "Pokémon" only ever appeared in one code comment; the app's tooltips already use plain "Pokemon" (the original spelling). No displayed string altered. Preserve é going forward.
val is not declared warning is harmless (the changeStr(string val) signal declares val; Qt 6 just deprecates injected handler params).

Session 13c — Build slowdown + dllimport spam: trim over-includes

Session 13 had included the entire expanded tree into savefileexpanded.h (included almost everywhere), dragging the heavy area sub-tree (whose .cpps pull mapdbentry.h/db) into nearly every TU → ballooned compile time; full rebuilds surfaced every pre-existing dllimport warning at once. Fix: include only the branches QML traverses — area.h → just areageneral.h, world.h → just worldother.h, dropped daycare/hof/rival from savefileexpanded.h; untraversed types back to forward-decl + Q_DECLARE_OPAQUE_POINTER. Added -Wno-ignored-attributes to root CMake to silence the harmless dllimport warning. Chain still works (verified no traversed type is opaque).

Tooling note: the bash sandbox mount lags behind the editor's writes and shows false truncation / stale content. The Read/Edit/Write file tools (and PowerShell) are the source of truth.

Session 13b — Chain works; pokemon types de-opaqued; number-box widths; About guard

Rebuild confirmed the chain fix — data flows into the UI. Then:

Pokémon box click didn't open details — getBoxMon()/getPartyMon() (Q_INVOKABLE returning PokemonBox*/PokemonParty*) feed PokemonDetails.qml's typed property PokemonBox boxData, which only works with real QObject pointers. Those pokemon-storage types were still opaque (s13 kept them opaque) → de-opaqued them + included their headers at the return/property sites. Nothing is opaque anymore in the savefile chain.
Number boxes too narrow (playtime, item counts) — width: 2 * font.pixelSize ignored the TextField padding → 2 * font.pixelSize + leftPadding + rightPadding.
About.qml:31 width of null — added parent ? parent.width : 0 guard.

Session 13 — REAL root cause of dataExpanded = undefined: Q_DECLARE_OPAQUE_POINTER

After the s12 rebuild the whole brg.file.data.dataExpanded.* chain was still undefined, disproving the s10–12 truncation/qRegisterMetaType theory (binary was current and registered).

Root cause: Q_DECLARE_OPAQUE_POINTER(T*) forces QtPrivate::IsPointerToTypeDerivedFromQObject<T*> = false, so Qt stores those QObject pointers as opaque non-QObject values and QML reads their sub-properties as undefined — and neither qRegisterMetaType nor qmlRegisterUncreatableType overrides it. Proof (natural experiment): brg.file worked (FileManagement fully #included in bridge.h, never opaque) while everything opaque-declared beneath it failed. This also produced the "Connections: no signal matches" flood (the targets were just undefined).

Fix: removed the opaque decls for the traversed QObject chain types and added full #includes down the chain so Qt sees them as real QObject pointers. See ../reference/qt-patterns.md → "`Q_DECLARE_OPAQUE_POINTER` breaks the QML property chain".