[skerker]

Summary

Fixes #6078.

On Windows, docking a floated panadapter lost the D3D11 device (NVIDIA: driver error and TDR reset; Intel: crash in the driver) when the pan's retained waterfall image was under 64 rows. The waterfall resize transaction in SpectrumWidget::renderGpuFrame() deleted the textures it replaced at once, in a frame that had already queued uploads into them. This PR releases them after the frame.

Root cause. In the first paint after a pan changes top-level window:

1. SpectrumWidget::render() calls initialize(cb) and then renderGpuFrame(cb, …) with the same command buffer.
2. initialize() → initWaterfallPipeline() sizes the waterfall textures from the retained image, raised to at least 64 × 64, and queues the initial uploads into them.
3. renderGpuFrame() compares the image with the texture. When the image is under the floor (measured: image 764 × 59, texture 764 × 64) they differ, so the resize transaction runs in that same frame, creates textures at the image size, and deletes the old ones.
4. The uploads queued in step 2 still name the deleted textures. Qt's D3D11 backend keeps a raw ID3D11Resource *dst in each queued upload command and has no deferred release (qrhid3d11_p.h, the same in Qt 6.8.3 and 6.12), so they execute at frame end against released objects.

With a retained image of 64 rows or more the texture is built at the image size, step 3 does not run in that frame, and nothing is released. That boundary is measured: 67 rows survives, the 64-row floor fails (table in #6078).

The change (one file, src/gui/SpectrumWidget.cpp):

• The five replaced resize resources (color texture, row-frequency texture, supplemental texture and two shader-resource bindings) use QRhiResource::deleteLater() so queued uploads finish before destruction. The debug resize line records the texture dimensions.
• releaseWaterfallFramePipelineResources() owns the five-resource cleanup list for both the row-texture resize fallback and idle teardown. It clears member pointers with std::exchange and uses deleteLater(), which defers during a frame and deletes immediately outside one. Both call sites share this lifetime rule without duplicating the resource list.

The 64 × 64 floor in initWaterfallPipeline() is unchanged, so the size mismatch still occurs; it is the in-frame release that was unsafe. These are the only two places in renderGpuFrame() that release resources with commands already recorded on the frame's command buffer. The in-place setPixelSize() + create() re-sizes of the overlay, frequency-overlay, background, 3DSS and FFT-column textures also release native textures, but they run before the frame's own batch is recorded, and initialize() queues nothing into those textures.

Containing an unrelated device loss after it has happened is outside this change. The operator explicitly chose to retain Fixes #6078 and close that issue with this trigger fix.

Possibly the same bug elsewhere. The attached logs of #5990 / #5935 (v26.9.5 and v26.8.4, AMD Radeon HD 7000, right-click → Pop out) end, in 5 of 7 files, inside the re-init after the pop-out, with waterfall row-frequency pipeline created 1624 x 64: the retained image's width and the 64-row floor. The logs of #5915 (Intel Arc, leaving Minimal Mode; closed by #5916) show waterfall row-frequency pipeline created 64 x 198, the floor on the width. That is a signature match, not a reproduction: those logs do not carry the image size. The reporter of #5990 can confirm on a build with this change. Refs #5990.

Constitution principle honored

Principle XI — Fixes Are Demonstrated. The before/after below is the implementer's evidence; the maintainer's reproduction is four steps (in #6078's body) on any Windows build.

Proof

Proof shape: automation bridge only. A GPU resource-lifetime fault has no socket-free test seam; no ctest is added.

Author evidence at 7021c4e0: one Windows 11 laptop with NVIDIA and Intel adapters. The Windows runs below were supplied by @skerker; they have not been rerun at the landing head. Independent macOS validation of the final head follows the author evidence. Linux rendering smoke has not been run.

All author runs: Windows 11 Pro 10.0.26200, x64, i9-14900HX, RTX 5060 Laptop GPU (driver 32.0.15.8205) + Intel UHD Graphics (32.0.101.7076), FLEX-8400 RX only (AETHER_AUTOMATION_NO_TX=1), an isolated copy of my settings (SpectrumSplitRatio 0.787, default floating size → 59-row floated waterfall, read with get rhi). Script: invoke "pan 0/panFloatToggle" / invoke "pan 1/panFloatToggle" alternately, 1.2 s apart.

Build	Adapter	Result
main 38792429 (Qt 6.12) and v26.9.5 portable (Qt 6.8.3)	NVIDIA	device lost on the first pan's 1st or 2nd dock, every run (#6078)
main 38792429	Intel (QT_D3D_ADAPTER_INDEX=1)	crash in igd10um64xe.DLL under d3d11!…UpdateSubresource ← QRhi::endOffscreenFrame, on the 1st dock
this PR, head 7021c4e0 (bridge ping SHA)	NVIDIA	survived 160 clicks: 80 inits at 764 x 64, 80 764 x 64 -> 764 x 59 lines, no 0x887a0005 line, no driver event, no crash
this PR, head 7021c4e0	Intel (get rhi: Intel(R) UHD Graphics)	survived 160 clicks: 80 inits at 764 x 64, 80 764 x 64 -> 764 x 59 lines, no crash
this PR, head 7021c4e0, AETHER_AUTOMATION_FORCE_WF_FRAME_RESIZE_FAILURE=1	NVIDIA	survived 160 clicks; the fallback branch ran 202 times (using legacy waterfall pipeline: row texture resize transaction failed), no loss, no driver event, no crash. Main was not run with this switch, so this shows the changed branch is safe, not that it failed before.

At the first commit 1a2d0545 the same script survived 160 clicks on each adapter: 80 inits at 764 x 64, 80 764 x 64 -> 764 x 59 lines, no 0x887a0005 line, no driver event, no crash.

The fix acting, from the after-run's log (the dock that lost the device in every earlier run):

11:14:04.248  invoke click on pan 0/panFloatToggle
11:14:04.397  SpectrumWidget: QRhi initialized, backend: D3D11
11:14:04.397  SpectrumWidget: waterfall row-frequency pipeline created 764 x 64
11:14:04.453  SpectrumWidget: waterfall texture resized 764 x 64 -> 764 x 59 (old resources released after the frame)
11:14:04.475  SpectrumWidget: QRhi initialized, backend: D3D11
11:14:04.475  SpectrumWidget: waterfall row-frequency pipeline created 764 x 64
11:14:04.533  SpectrumWidget: waterfall texture resized 764 x 64 -> 764 x 59 (old resources released after the frame)
11:14:04.597  SpectrumWidget: waterfall texture resized 764 x 59 -> 1192 x 186 (old resources released after the frame)

Evidence bundle: head 7021c4e0 · whole logs and driver records for the three head runs, one before pair on unchanged main (NVIDIA loss, Intel crash with the debugger stack), the bridge scripts, both commits as a patch · aethersdr-pr6093-waterfall-resize-release-2026-10-02.zip (sha256 33a5822a8728fdfd89bb794d6b43190ec36aad77370198a860928be41f2467fb)

Landing validation at 150059be0321c35d132a271d0d5917c1b118633e

Upstream source base a7eccda4 was merged without rewriting the contributor commits. The final PR diff remains one source file. The cleanup and clean integration commits are signed and GitHub verified. Main advanced twice during the pass: split-QSY handling (5766bb13) and WNB startup replay removal (a7eccda4). Both were incorporated before the final-head build and runtime proof.

• macOS, Apple M2 Ultra, Qt 6.12.0, GPU spectrum enabled: cmake --build build -j22 --target AetherSDR spectrum_preview_logic_test waterfall_history_buffer_test waterfall_time_markers_test waterfall_rate_test passed.
• Headless socket-free CTest selection: spectrum_preview_logic_test, waterfall_history_buffer_test, waterfall_time_markers_test, waterfall_rate_test — 4/4 passed with QT_QPA_PLATFORM=offscreen.
• Native Metal bridge smoke against isolated DEMO-0001, TX disabled: 40 alternating float/dock clicks across two panadapters and four window resizes passed. The log contains 51 transitions from the 64-row texture floor to a smaller retained waterfall. Bridge state remained healthy (rendererFailed: false, texture/image dimensions matched after settling).
• A separate native Metal instance with AETHER_AUTOMATION_FORCE_WF_FRAME_RESIZE_FAILURE=1: 40 float/dock clicks and four window resizes passed, including 61 forced fallback events and 51 transitions from the 64-row floor. This exercises the shared cleanup helper during a recording frame. Both screenshots showed spectrum/waterfall rendering; both instances disconnected and exited cleanly.
• Engine-boundary, touchpoint-manifest, capability-record and command-plane static gates passed; merge-base and head findings are identical. git diff --check passed.
• Fresh hosted CI runs on this head were created at 20:20:50Z on October 2, 2026, after the current main tip. All five PR checks subsequently passed at this head.

The four CTests are adjacent logic coverage, not a reproduction of the GPU lifetime fault. The author’s Windows before/after table is the regression reproduction; the final-head Metal smoke independently covers normal, forced-fallback, two-panadapter, resize and idle-teardown paths. No live hardware was used during landing validation.

After the final-head CI runs were created, main gained 889bbe8d23735b60ce08e8f7bd44d1a45b4b34ea (20:35:38Z), docs(bridge): automation identity also picks the radio's restore. Principle VIII. (#6071). Its only changed file is docs/automation-bridge.md; no src/, tests/, tools/, workflow or CMake changes. The pr-land docs-only freshness exemption applies. Merge-tree is clean and preserves that documentation update.

Not tested

• Windows runtime smoke on the final landing head; the Windows table above is evidence at the contributor’s 7021c4e0 head. Hosted Windows CI is separate build evidence.
• Linux rendering smoke, other machines/adapters, and the AMD card of [bug] panadapter righ click choose Pop Out crash AESDR #5990.
• Arbitrary device-loss containment and recovery after a loss caused elsewhere.
• Removing the 64 × 64 floor; it is unchanged.

Test plan

• Final-head macOS build with GPU spectrum enabled.
• Four focused socket-free tests passed at the final head.
• Final-head normal and forced-fallback Metal bridge smoke, isolated demo, TX disabled.
• Author-supplied Windows reproduction and before/after evidence at 7021c4e0.
• Fresh hosted CI at the final head — all five checks passed.

Checklist

• Commits are signed (docs/COMMIT-SIGNING.md)
• No new flat-key AppSettings calls — n/a, no settings touched
• Code is clean-room — not decompiled, disassembled, or reverse-engineered from a proprietary binary (Principle IV); Qt's own shipped header is the only outside source read
• All meter UI uses MeterSmoother — n/a, no meter UI
• Documentation updated if user-visible behavior changed — n/a, no user-visible behavior change; CHANGELOG.md untouched
• Security-sensitive changes reference a GHSA if applicable — n/a

Source integration base: upstream main a7eccda4ec9341485982091734301b8b116f21ca incorporated in the signed landing commit; no unrelated PR diff.

Original fix and Windows evidence by @skerker, authored with Claude Code. Landing cleanup and macOS validation by Codex under @jensenpat.

Post-merge verification: squash commit 712bd84994f55b8c484e3cb95a9e02b469242508 has the expected tree and a valid GitHub signature. Platform CI and Full Suite passed on that exact main commit. C++ CodeQL and System Libraries Canary were still running at 21:57Z. #6078 closed with the trigger fix, as requested.

Generated with OpenAI Codex (GPT-6 Astra)

[aethersdr-agent]

Issue fit

Yes, for the defect the issue now names. I walked the root-cause chain in the head checkout and it holds end to end: render() calls initialize(cb) then renderGpuFrame(cb, …) with the same command buffer (SpectrumWidget.cpp:15135-15165); initialize() queues batch->uploadTexture() into exactly m_wfGpuTex, m_wfSupplementalGpuTex and m_wfFrameTex and commits with cb->resourceUpdate(batch) (:13480-13513); initWaterfallPipeline() floors the texture at 64 × 64 for a shorter retained image (:12566-12569); and the resize transaction at :13640 is reached in that same frame, since desiredWidth/Height come straight from m_waterfall with no floor. The three resources the old code deleted in-frame are precisely the three initialize() uploaded into. Nothing in the diff is ornamental.

I also closed the issue's one open "Not known" — "which statement mishandles the rows between the image height and the 64-row texture". It is not an upload-size bug: prepareWaterfallFrameUpload() sizes the buffer max(visibleHeight, m_wfGpuTexH) and pads the floor rows with the current center/bandwidth (:13318-13360), and the color/supplemental uploads are legal partial uploads into a taller texture. That is consistent with this PR's account — the mismatch is benign, the in-frame release was not.

The mechanism claim I could check: qrhi_p.h:178-183 in the Qt 6.11.2 kit here is

void addDeleteLater(QRhiResource *res)
{ if (inFrame) pendingDeleteResources.insert(res); else delete res; }

so deleteLater() defers only while a frame is recording, deletes immediately otherwise, and pendingDeleteResources is a QSet (double-deferral of one pointer is harmless). What I could not check here: that the set is flushed after the backend's endOffscreenFrame (that is in qrhi.cpp, not shipped in this kit), and the PR's specific statement that the D3D11 backend alone lacks a deferred-release queue. Those rest on your 160-click before/after runs, not on anything I verified. Everything below is reasoned from source; I have no build and no GUI here.

What stays unfixed: #6078's Expected has two halves — "pop-out and dock-back keep working" (this PR) and "if the D3D11 device is lost, the pan reports a render failure instead of the process crashing". The thread's triage comment is entirely about that second half (device-loss latch, gating update() drivers, the VCRUNTIME140/QBackingStoreDefaultCompositor crash path with no AetherSDR frame in it), and the PR correctly scopes it out. Fixes #6078 will auto-close the issue carrying that work. Worth either demoting to Refs #6078 or filing the containment follow-up before merge — maintainer's call, not a blocker on this code. Nothing in the thread rules against fixing the trigger instead; the one alternative the triage floated (single-init dock) was measured and did not help.

Scope

File / group	What it changes	Claimed by title/body	Verdict
src/gui/SpectrumWidget.cpp:13709-13718	releaseAfterFrame lambda + comment; new qDebug resize line	Yes (commit 1a2d0545, the log line is quoted in the Proof section)	In scope
src/gui/SpectrumWidget.cpp:13731-13737	three in-frame deletes → deleteLater()	Yes	In scope
src/gui/SpectrumWidget.cpp:13738-13747	fallback branch: five std::exchange + deferred release before the helper	Yes (commit 7021c4e0)	In scope
src/gui/SpectrumWidget.cpp:13753-13754	oldLegacySrb / oldColorTexture → deferred	Yes	In scope

One file, +27/−5, no new public surface, no settings or protocol touched, no guard or early return removed, CHANGELOG.md untouched (correct — don't add one). The checklist's "no unrelated files" self-certification is accurate. No sibling implementation left behind: grep finds no other delete between renderGpuFrame's start and releaseResources(), and the five setPixelSize()+create() recreates (:14005, :14011, :14035, :14783, :14851) hit textures initialize()'s batch never names, each one preceding its own upload later in the same frame (:14360, :14382, :14794, :14882) — so the body's "these are the only two places" claim checks out. No test changes, no sockets, no fake peers; a GPU resource-lifetime fault has no socket-free seam and the "bridge-only proof" framing is the honest one. Not a preference change: no default, no UI element, no workflow moved.

Blockers

None.

Nits (non-blocking)

• The five std::exchange lines duplicate releaseWaterfallFramePipelineResources()'s body (inline comment at :13741). Because addDeleteLater already falls back to plain delete when no frame is recording, the helper itself could use deleteLater() and both call sites — this one and releaseResources() at :15171 — would be correct with no duplication. As written, anyone who adds a sixth resource to the helper silently reintroduces the in-frame delete on this path.
• Fixes #6078 vs. the containment half of that issue's Expected, as above.
• The new qDebug() line fires once per waterfall resize, so a dock burst or a resize settle writes several; it is default-category like its siblings (:12757, :13405), which is what makes your proof log readable in a bundle, so I would not change the category — just be aware it competes for the bundle tail with whatever else is happening during a reparent.

What I tried to break

• Double release across the two branches. oldFrameSrb/oldFrameTexture/oldSupplementalTexture are the same objects the fallback branch std::exchanges out of the members. Each path releases them exactly once, and the helper's delete nullptr after the exchange is a no-op; pendingDeleteResources being a QSet makes even a repeat insert harmless. No double free.
• Resources still deleted inside the frame. On every failure path (!legacyResizeReady, row-texture create() failure, frameSrb->create() failure) the unique_ptrs destruct immediately — but those are the textures/SRBs created this frame, never enqueued on the batch and never bound in a recorded pass. Safe.
• A stale SRB surviving the resize. If m_wfPipelineMode == RowFrequencyFrames but rowPipelineWasActive were false, m_wfFrameSrb would keep sampling the color texture that was just released, and the draw gate at :14944-14951 would still use it. I could not reach that state: m_wfSupplementalGpuTex is only assigned at :12749/:13731 (alongside the rest) or nulled by the helper, which also forces m_wfPipelineMode = Legacy and m_wfFrameTexReady = false. The deferral makes this strictly safer than before, not worse.
• The live-resize preview path. render() can also call renderGpuFrame(cb, m_resizePresentationSize, true) during the settle timer — same function, same transaction, so it is covered by the same change; the preview flag does not bypass the resize block.
• Compile portability. std::exchange is covered (<utility> at line 94), and the hunk is inside #ifdef AETHER_GPU_SPECTRUM. All 5 checks are green on 7021c4e0, but note CMakeLists.txt:640-647 silently sets AETHER_GPU_SPECTRUM OFF when Qt6GuiPrivate is absent, so green CI is only evidence this compiled wherever that module was found — the macOS/Linux float/dock smoke you list as untested is still the real gap, and deleteLater() is backend-agnostic so I see no reason to expect a platform difference.

Recommendation

Approve with nits. The diff is small, the root cause is traceable in the source rather than asserted, the before/after runs cover both adapters plus the forced-failure branch, and I could not construct a double free, a leak, or a stale binding out of either hunk. The two things I would like settled before merge are bookkeeping, not code: collapse the duplicated release list into the helper (or say why not), and decide whether this PR should close #6078 given the device-loss containment work the thread asks for is explicitly not in it. Thanks for the measurement table — the 64/67-row boundary is what made this reviewable.

---

_{🤖 aethersdr-agent · cost: $4.9145 · model: claude-opus-5}

[aethersdr-agent]

Two line references in my review above are wrong: prepareWaterfallFrameUpload() is at src/gui/SpectrumWidget.cpp:5807-5851 (not :13318-13360), and the 64 × 64 floor is at :12568-12571 (not :12566-12569). The findings are unchanged.

---

_{🤖 aethersdr-agent · cost: $5.5679 · model: claude-opus-5}

[jensenpat]

Landing disposition at 85943b0482f0aad5a8bdfcb1bb6558d2b17cf07f: the duplicated resource list is now centralized in the existing helper and proved by the final-head Metal forced-fallback smoke (61 fallback events). The two source-lifetime caveats are answered with Qt v6.12.0 primary-source links in the inline thread. The corrected upload/floor line references were included in the manual source audit. The resize debug log is retained as reviewed and used to verify the floor transitions.

The operator explicitly chose keep Fixes #6078; close with this trigger fix. The PR body retains that keyword and states that arbitrary device-loss containment is outside the change. The body now distinguishes author-supplied Windows evidence at 7021c4e0 from final-head macOS build, four focused tests, and normal/forced-fallback Metal smoke. Fresh CI is running. The automated code-review skill was unavailable in this session; the adversarial pass was performed manually.

[jensenpat]

Approve exact head 85943b0482f0aad5a8bdfcb1bb6558d2b17cf07f after manual adversarial review, remediation and runtime proof.

• Fixed the duplicated row-pipeline cleanup list in SpectrumWidget.cpp:12538: the existing helper uses null-guarded deleteLater() and std::exchange, shared by in-frame fallback and idle teardown. The final-head GPU-enabled macOS build passed, four focused headless socket-free CTests passed, and the forced-fallback Metal bridge run exercised 61 fallback events while rendering remained healthy.
• Closed the flush-order/raw-pointer caveats with public Qt v6.12.0 source: backend endOffscreenFrame() precedes draining pending resources; the D3D11 queued update structure retains raw destination pointers. Both inline threads have evidence replies and are resolved; no blocking reviews needed dismissal.
• Final-head isolated DEMO-0001, TX-disabled Metal smoke: normal and forced-fallback runs each passed 40 alternating two-panadapter float/dock clicks and four window resizes. Logs captured 54 normal and 51 forced-fallback transitions from the 64-row floor to a smaller image. Screenshots showed spectrum/waterfall; both instances disconnected and exited cleanly. The shared helper's null/idle paths and failure branches were also audited in source.
• Operator decision: retain Fixes #6078 and close the issue with this trigger fix. Arbitrary device-loss containment is outside the change and is stated in the body. Resize logging is retained as reviewed. No settings, radio protocol, defaults, workflows, changelog or architecture ratchets changed; all four local static-gate findings match the merge base.

Author-supplied Windows before/after proof is at 7021c4e0; Windows runtime smoke at this final head and Linux rendering smoke were not rerun. Hosted Windows/Linux build checks are separate CI evidence. The automated code-review skill was unavailable; this was a manual adversarial pass. Fresh hosted CI and current-main freshness are verified separately before arming the merge.

[jensenpat]

Approve exact head 150059be0321c35d132a271d0d5917c1b118633e after manual adversarial review, remediation and runtime proof.

• Fixed the duplicated row-pipeline cleanup list in SpectrumWidget.cpp:12538: the existing helper uses null-guarded deleteLater() and std::exchange, shared by in-frame fallback and idle teardown. The final-head GPU-enabled macOS build passed, four focused headless socket-free CTests passed, and the forced-fallback Metal bridge run exercised 61 fallback events while rendering remained healthy.
• Closed the flush-order/raw-pointer caveats with public Qt v6.12.0 source: backend endOffscreenFrame() precedes draining pending resources; the D3D11 queued update structure retains raw destination pointers. Both inline threads have evidence replies and are resolved; no blocking reviews needed dismissal.
• Final-head isolated DEMO-0001, TX-disabled Metal smoke: normal and forced-fallback runs each passed 40 alternating two-panadapter float/dock clicks and four window resizes. Logs captured 51 normal and 51 forced-fallback transitions from the 64-row floor to a smaller image. Screenshots showed spectrum/waterfall; both instances disconnected and exited cleanly. The shared helper's null/idle paths and failure branches were also audited in source.
• Operator decision: retain Fixes #6078 and close the issue with this trigger fix. Arbitrary device-loss containment is outside the change and is stated in the body. Resize logging is retained as reviewed. No settings, radio protocol, defaults, workflows, changelog or architecture ratchets changed; all four local static-gate findings match the merge base.

Author-supplied Windows before/after proof is at 7021c4e0; Windows runtime smoke at this final head and Linux rendering smoke were not rerun. Hosted Windows/Linux build checks are separate CI evidence. The automated code-review skill was unavailable; this was a manual adversarial pass. Fresh hosted CI and current-main freshness are verified separately before arming the merge.

Revalidated after cleanly incorporating the two intervening main commits (5766bb13, a7eccda4). Earliest final-head CI run creation is 2026-10-02T20:20:50Z, later than the incorporated source base’s 2026-10-02T20:15:54Z commit time. Required checks are pending. Auto-merge is confirmed armed at this head. An initial request with an explicit contributor --author-email override was rejected with a misleading verified-email message; removing that override succeeded. No operator account change was needed. All required checks must pass before merge.

After the final-head CI runs were created, main gained 889bbe8d23735b60ce08e8f7bd44d1a45b4b34ea (20:35:38Z), docs(bridge): automation identity also picks the radio's restore. Principle VIII. (#6071). Its only changed file is docs/automation-bridge.md; no src/, tests/, tools/, workflow or CMake changes. The pr-land docs-only freshness exemption applies. Merge-tree is clean and preserves that documentation update.

[skerker]

Linux rendering smoke at 7021c4e0 (before the maintainer commits): no failures. The forced-fallback path changed in 85943b04 was not run on Linux.

Setup: Ubuntu 24.04, X11, Qt 6.12.0 (pinned kit from scripts/setup/setup-qt.sh), OpenGL on an NVIDIA RTX 5060 Laptop GPU (driver 595.91.07). FLEX-8400, receive only (AETHER_AUTOMATION_NO_TX=1). Isolated copy of my settings with SpectrumSplitRatio 0.787, driven by the same bridge script as the Windows runs. Bridge ping reported sha 7021c4e0 in both runs.

Run	What was driven	Result	waterfall texture resized … released after the frame	QRhi failure / legacy fallback
Float/dock	160 alternating float/dock clicks on two pans	160 of 160, app alive	285, of which 40 × 764 x 64 -> 764 x 59	0 / 0
Window resize	10 maximize/restore, then two height sweeps (1348 ↔ 900 px)	app alive, both waterfalls kept painting	143 (waterfall heights 38–91 rows)	0 / 0

• The floated waterfall on pan 0 was 59 rows, so the 64-row floor case from Windows: docking a floated panadapter whose waterfall is under 64 rows loses the D3D11 device and crashes #6078 was reached on Linux too: the log has 40 waterfall row-frequency pipeline created 764 x 64 lines and 40 resizes from 764 x 64 to 764 x 59.
• A float or dock re-initialises the pan on Linux as well (162 QRhi initialized lines for 160 clicks).
• get rhi reported rendererFailed: false with texture size equal to image size at every reading.

Between 7021c4e0 and the current head 150059be, src/gui/SpectrumWidget.cpp differs in three places: the settings clean-up from #6070, the shared release helper, and the fallback branch. The resize transaction these runs exercised is the same code.

— authored by agent (Claude Code) on behalf of @skerker