Merge: fix reload-in-Design mis-tagging Arrange tracks into Design

src/bank_panel.cpp

@@ -230,15 +230,26 @@ struct PanelState {
     //
     // Each timer tick diffs the live track+item GUID set against the previous tick to
     // auto-tag content created SINCE the last tick into the then-active mode. The
-    // baseline carries the first-poll-after-open guard so pre-existing content is never
-    // mass-tagged (it stays Arrange). `lastProject` detects a project switch so the
-    // baseline re-arms per project (a switch never diffs across two projects). Both live
-    // for the extension's lifetime alongside the session, independent of panel open/close
-    // — detection must run whether or not the dock is visible (content is created in the
-    // arrange, not the panel).
+    // baseline carries the first-poll-after-open guard (GuidBaseline self-arms on its
+    // first observe()) so pre-existing content is never mass-tagged (it stays Arrange).
+    //
+    // Project-load re-arm is driven by persist's AUTHORITATIVE load lifecycle, NOT by a
+    // pointer compare here. main.cpp calls bankPanelNotifyProjectLoaded() on the exact
+    // tick persist restores a project's membership + active mode (the same tick it
+    // reapplies the active mode); that sets reloadPending so the NEXT detect tick this
+    // same tick re-baselines against the fully-loaded set and reports nothing new. This
+    // replaces the former `proj != lastProject` re-arm, which used a WEAKER signal than
+    // persist (pointer-only vs persist's GUID-primary identity) and so missed a load onto
+    // a RECYCLED ReaProject* address — the just-loaded project's pre-existing tracks then
+    // diffed against the previous project's stale baseline and were mass-tagged into the
+    // active mode (the reload-mis-tag bug). Coordinating with persist's signal makes the
+    // two identity checks agree by construction.
+    //
+    // Lives for the extension's lifetime alongside the session, independent of panel
+    // open/close — detection must run whether or not the dock is visible (content is
+    // created in the arrange, not the panel).
     GuidBaseline contentBaseline;
-    ReaProject*  lastProject = nullptr;
-    bool         sawProject = false;  // false until the first detect tick sees a project
+    bool         reloadPending = false;  // set by bankPanelNotifyProjectLoaded; drained next detect tick
 };
 
 PanelState g_panel;
@@ -747,14 +758,19 @@ void detectNewContent() {
 
     ReaProject* proj = EnumProjects(-1, nullptr, 0);
 
-    // Project switch (or first ever tick) re-arms the first-poll guard so we never diff
-    // across two projects. GUID-address recycling is bounded here: a missed reset can at
-    // worst re-baseline against the wrong project for one tick; the identity-of-record
-    // (persist's minted project GUID) governs the bank/model reload, not this detector.
-    if (!g_panel.sawProject || proj != g_panel.lastProject) {
+    // A project (re)load re-arms the first-poll guard so we never diff across two
+    // projects. The signal is persist's — main.cpp calls bankPanelNotifyProjectLoaded()
+    // on the tick persist restores the project's membership + active mode, which sets
+    // reloadPending. Draining it here re-baselines against the fully-loaded set (that
+    // same tick's reapply-active-mode enumerated those tracks, so they are present),
+    // and the observe() below returns nothing new — pre-existing untagged tracks stay
+    // Arrange. GuidBaseline self-arms on its first observe() for the very first tick, so
+    // no separate first-tick handling is needed here. Using persist's GUID-primary load
+    // signal (not a local pointer compare) is what fixes the reload-mis-tag: the two
+    // identity checks can no longer diverge on a recycled ReaProject* address.
+    if (g_panel.reloadPending) {
         g_panel.contentBaseline.reset();
-        g_panel.lastProject = proj;
-        g_panel.sawProject = true;
+        g_panel.reloadPending = false;
     }
 
     std::set<std::string> live;
@@ -1212,6 +1228,18 @@ std::vector<std::string> bankPanelSelectedSampleIds() {
     return ids;
 }
 
+void bankPanelNotifyProjectLoaded() {
+    // Persist restored a project's membership + active mode this tick (main.cpp calls
+    // this from the same consumeLoadSignal() branch that reapplies the active mode).
+    // Arm the new-content detector to re-baseline on its next tick so the just-loaded
+    // project's pre-existing content is treated as the baseline (nothing new) rather
+    // than diffed against the previous project and mass-tagged into the active mode.
+    // A flag (not an inline reset) because detectNewContent owns the baseline and runs
+    // later in the SAME OnTimer tick — it drains this and re-baselines against the live
+    // set in one place, keeping the reset and the observe() adjacent and ordered.
+    g_panel.reloadPending = true;
+}
+
 void bankPanelRefresh() {
     // New-content auto-tag detection runs EVERY tick regardless of panel open/close:
     // tracks/items are created in the arrange view, not the panel, so detection must

src/bank_panel.h

@@ -50,6 +50,17 @@ std::vector<std::string> bankPanelSelectedSampleIds();
 // reflected without the panel diffing the bank itself.
 void bankPanelRefresh();
 
+// Notifies the panel that persist just (re)loaded a project's view model (membership +
+// active mode). main.cpp calls this on the exact tick it drains persist's load signal
+// and reapplies the active mode. It re-arms the new-content detector so the just-loaded
+// project's PRE-EXISTING content is taken as the baseline (reported as nothing new),
+// never diffed against the previously-open project and mass-tagged into the active mode.
+// This coordinates the detector's project-identity signal with persist's authoritative
+// (GUID-primary) one — the two can no longer diverge on a recycled ReaProject* address,
+// which is what caused a project opened in Design to mis-tag its Arrange tracks. READ/
+// arm of panel state only; no project or bank mutation.
+void bankPanelNotifyProjectLoaded();
+
 // The panel's current tail-mode setting (mode + Manual length), read by the plain
 // CAPTURE_ITEM / CAPTURE_TRACK actions when building a CaptureRequest so a capture
 // applies whatever the panel toggle is set to. Default None (exact bounds) — a

src/main.cpp

@@ -216,8 +216,16 @@ static void OnTimer()
     // project saved in Design mode parks the Arrange tracks automatically, no manual
     // toggle. Fires exactly once per load (consumeLoadSignal clears it); idle ticks
     // skip it. proj = nullptr -> REAPER's active project (the one poll just loaded).
-    if (g_session.consumeLoadSignal())
+    //
+    // The SAME signal re-arms the bank panel's new-content detector: a load must
+    // re-baseline the detector against the just-loaded project's content so its
+    // pre-existing tracks are never mis-detected as "new" and mass-tagged into the
+    // active mode (the reload-mis-tag bug). Notify BEFORE the reapply so the detector's
+    // re-arm and the model restore ride the one authoritative load event.
+    if (g_session.consumeLoadSignal()) {
+        reasampler::bankPanelNotifyProjectLoaded();
         reasampler::applyMode(g_session.view(), g_session.view().activeModeId(), nullptr);
+    }
 
     // Reflect a live bank change (capture / project load) in the docked grid.
     // Cheap when the bank is unchanged (a fingerprint compare); repaints only on

tests/test_guid_diff.cpp

@@ -124,6 +124,61 @@ static void testResetReBaselines() {
     CHECK(p2new.size() == 1 && has(p2new, "{W}"));
 }
 
+// -- 6. Reload-mis-tag regression: a project LOAD must re-baseline before the first
+//       post-load observe, so the newly-loaded project's PRE-EXISTING content is never
+//       reported as new. This locks the exact failure behind the reload-mis-tag bug:
+//       the detector used to re-arm on a `proj != lastProject` pointer compare, which a
+//       recycled ReaProject* address defeats; the previous project's stale baseline then
+//       reported the whole just-loaded project as new content and it got mass-tagged into
+//       the active mode. The fix routes the re-arm through persist's authoritative load
+//       signal (bankPanelNotifyProjectLoaded -> reset()), modeled here as: on a load,
+//       reset() runs BEFORE the first observe of the new project's set.
+//
+//       The seam under test is GuidBaseline; the shell wiring (main.cpp notify ->
+//       bank_panel reset()) is DAW-verified, but the load-then-observe DECISION lives
+//       here and is what the bug got wrong.
+static void testReloadReBaselinesBeforeFirstObserve() {
+    // Project A is open and settled: its content is the baseline, steady state reports
+    // nothing new. This is the "extension already running against project A" precondition
+    // the bug needs (a NON-empty stale baseline to mis-diff the next project against).
+    GuidBaseline b;
+    b.observe({"{A1}", "{A2}"});                 // A baseline (first-poll guard)
+    CHECK(b.observe({"{A1}", "{A2}"}).empty());  // steady: nothing new
+    CHECK(b.primed());
+
+    // Daniel opens project B (saved in Design). B's pre-existing tracks are an ENTIRELY
+    // different GUID set from A. persist raises its load signal; the fix calls reset()
+    // (via bankPanelNotifyProjectLoaded) BEFORE the first post-load observe.
+    b.reset();
+    auto afterLoad = b.observe({"{B1}", "{B2}", "{B3}"});
+    // The load must tag NOTHING: B's pre-existing content is the baseline, not "new".
+    // Untagged/Arrange leaves stay Arrange; nothing is mass-tagged into Design.
+    CHECK(afterLoad.empty());
+
+    // And genuine post-load creation in B is still detected (the fix must not deafen the
+    // detector — only suppress the pre-existing set at the load boundary).
+    auto createdInB = b.observe({"{B1}", "{B2}", "{B3}", "{B4}"});
+    CHECK(createdInB.size() == 1 && has(createdInB, "{B4}"));
+}
+
+// -- 6b. Negative control: WITHOUT the load re-baseline (the old pointer-miss path where
+//        reset() never fired), the just-loaded project's pre-existing content IS reported
+//        as new — i.e. it would be mass-tagged. This proves the assertion in test 6 is
+//        load-bearing (the reset() is what prevents the mis-tag), not self-affirming.
+static void testMissingReBaselineWouldMisTag() {
+    GuidBaseline b;
+    b.observe({"{A1}", "{A2}"});                 // A baseline
+    b.observe({"{A1}", "{A2}"});                 // settled against A
+
+    // Simulate the BUG: no reset() on the load (the pointer compare missed a recycled
+    // ReaProject*). The next observe diffs B's set against A's stale baseline.
+    auto misdetected = b.observe({"{B1}", "{B2}", "{B3}"});
+    // Every one of B's pre-existing tracks looks "new" — exactly the mass-tag that
+    // parked the Arrange tracks into Design on open. This is the failure the fix removes.
+    CHECK(misdetected.size() == 3);
+    CHECK(has(misdetected, "{B1}") && has(misdetected, "{B2}") && has(misdetected, "{B3}"));
+}
+
 int main() {
     testNewGuidsDifference();
     testNewGuidsIgnoresEmpty();
@@ -131,6 +186,8 @@ int main() {
     testBaselineIncremental();
     testBaselineDeletionReDetect();
     testResetReBaselines();
+    testReloadReBaselinesBeforeFirstObserve();
+    testMissingReBaselineWouldMisTag();
 
     if (g_fail == 0) std::printf("All tests passed.\n");
     return g_fail ? 1 : 0;