[on8st]

Update after review, efcbb06b and b75b8356 (full suite 611 of 611):

• Icom and the demo are not fixed. Their rows are dB too (IcomCivBackend.cpp:1730, SimSignalSource.cpp:137) but they declare no binsAbsolute, so NB Blank stays dead there. The predicate is the maintainer's to choose; the options are in the review reply.
• The rows are dBFS under a dBm label: the code now says RowKind::AbsoluteDb. Read "dBm row" below as "absolute dB row".
• The ring empties when the row kind changes, and its size is one constant shared by the widget and the test.

Summary

Refs #5678 (the HL2 parity rows). Refs #277 (the feature).

Read first.

• The fix has not been seen in the running app. The defect was observed there. The fix is covered by a unit test on the extracted decision.
• One design choice, yours to move: what the operator's threshold means on a dBm row. This PR reads it as one dB per slider step (kDbPerThresholdUnit = 100).
• It reaches three families. The switch is RadioCapabilities::panBinsAbsolute(), which HL2, ANAN and RTL-SDR declare. Only the HL2 was observed.
• The test ends in a source-text pin on SpectrumWidget.cpp. It names its claim; if you would rather not carry it, it is four checks at the end of one file.

On a Hermes-Lite 2 the waterfall's NB Blank is enabled, toggles, persists and does nothing.

Observed (the defect, before the change)

HL2 on a dummy load, receive only, 2026-10-01, on a local integration build over 25da98de (Qt 6.12.0). The floor was stepped up 35.0 to 36.8 dB with the LNA gain and the waterfall grabbed 2.00 s later; three cycles each. get display confirmed wfBlankerEnabled and the threshold on every leg.

NB Blank	n	floor step, dB	burst depth, px
Off	3	35.6 to 36.1	50, 50, 50
On, 1.05	3	36.0 to 36.8	50, 50, 50
On, 1.95	3	35.0 to 35.2	50, 50, 50

At 25 rows/s, 2.00 s is 50 rows, one px a row: the step is in the waterfall with no delay on 9 of 9 legs. On equals Off. There was no positive control: no working blanker was seen on any radio in that run.

Cause

SpectrumWidget::updateWaterfallRow replaced a row when

m_wfBlankerRingCount >= 8 && baseline > 0.0f && rowMean > baseline * m_wfBlankerThreshold

A Flex tile is a positive intensity (about 96..120), so that is a ratio test. A radio whose spectrum is computed on this host has no waterfall plane: RadioModel::onBackendSpectrumFrame hands the pan frame on as the row. That frame is dBm, every bin negative, so baseline > 0.0f is false on every row.

Dropping the guard is not the fix. With a baseline of -147 and the default 1.15, baseline * threshold is -169; every ordinary row is above that, and the waterfall would stand still.

The change

• src/gui/WaterfallImpulseBlanker.h (new, no Qt): decide(kind, historyRows, baseline, rowMean, threshold) returns whether the row is an impulse and what it adds to the ring.
• SpectrumWidget.cpp: the inline test and the two ring writes become one call. The row kind is m_panBinsAbsolute, the declared capability the widget already holds for the auto-floor gate. It is not guessed from the sign of the data.
• Tile rows are unchanged: same test, same 1.05 cap on what a rejected row adds to the ring.
• dBm rows: impulse when historyRows >= 8 and rowMean - baseline > (threshold - 1) * 100 dB. A rejected row enters the ring at most 5 dB over the baseline (the tile cap through the same mapping). A baseline that is not finite is never an impulse: one row mean of -inf in the ring would otherwise make every later row an impulse and write -inf back.

The design choice: what the threshold means in dB

The control is a slider 5..95 that becomes 1 + v/100, default 1.15. This PR reads it as one dB per slider step: 5 dB at the lowest setting, 15 dB at the default, 95 dB at the top.

On a tile, mean > baseline * t means a margin of (t - 1) times the tile's own level, about 96..120, so one slider step is about one intensity unit. A dBm row cannot use its own level, because the zero of a dBm scale is arbitrary. So the margin uses a fixed 100. One dB is taken as one intensity unit because the widget already colours both row kinds with one range width in the row's own unit (120 - 0.91 * gain). What one tile unit is worth in dB is not measured here.

law	at 1.05	at 1.15	at 1.95	why not
(t - 1) * 100 (this PR)	5 dB	15 dB	95 dB
10 * log10(t)	0.2 dB	0.6 dB	2.9 dB	Below the movement of an ordinary row's mean: a station 40 dB over the floor on 5 % of the bins lifts it by 2 dB, and the waterfall would hold each time it keys up
(t - 1) * abs(baseline)	5.6 to 7.4 dB	17 to 22 dB	105 to 141 dB	Moves with LNA gain (floor -111 to -148 dBm on this HL2)

Another law is one constant or one line in dbMargin.

Tests

waterfall_impulse_blanker_test (new, no Qt, 47 checks):

• dBm rows. A row 20 dB over a -147 dBm floor is replaced at the default; a row 0.5 dB over, at, or below the baseline is not. 7 rows of history is not enough, 8 is. The margin is 5, 15 and 95 dB. The same excess gives the same answer on baselines from -148 to +10 dBm.
• dBm rows through the widget's ring. A steady band at -60 dBm is never blanked in 400 rows. A single impulse row is replaced and moves the baseline by 5/32 dB. A 36 dB step that stays is held for 108 rows at 1.05 and 75 at 1.15, then released; at 1.95 it passes. A falling level is never blanked. A -inf row does not freeze the waterfall.
• Tile rows. Nine hand-computed cases, and 121 176 cases (history 0..33, baselines of both signs, six thresholds) that decide bit for bit as the original expression, 27 040 of them impulses.
• Source-text pin (four checks). The claim: SpectrumWidget.cpp takes the row kind from m_panBinsAbsolute, writes the helper's value to the ring, and keeps no inline ratio test. No behavioural seam reaches it: SpectrumWidget is compiled only into the application, and no test target links it. It shows how the call is written, not that the widget runs.

Red before green, re-run at this head. The helper is new, so the test cannot build on main; "the dBm branch inert" is main's behaviour.

mutation	failing checks of 47
the dBm branch inert (main)	11
the ratio test on dBm rows, guards dropped	14
10 * log10(t) as the margin	13
the rejected row uncapped	4
the finite guard dropped	3
the two kinds swapped at the call site	1 (the source-text pin)

Mutations were run with the test compiled on its own with Apple clang.

Full ctest on macOS (Qt 6.12.0, Ninja) at 63a73a35, which contains main at c7513e12: 100% tests passed out of 611.

Every script step of the Static checks job in static-checks.yml was run locally with stdin from /dev/null and is clean, the colour ratchet at +0. check_a11y on the two changed GUI files: no findings.

Not verified

• Not observed in the running app. The observation above is the defect, on a build without this change. No build carrying the fix has been launched against a radio.
• The observed stimulus is a sustained step, not an impulse. No real impulse on an HL2 waterfall has been looked at.
• ANAN and RTL-SDR take the same path; nothing was run on them. Icom and the demo backend leave the flag false and keep the tile test.
• The blanker firing on a Flex is read from the source and from Noise Blanker–Gated Waterfall Impulse Suppression #277, not observed here.
• macOS only. The sanitizer-configure job was not run locally.

Beside this on main

Nothing on main overlaps: the branch merges c7513e12 cleanly. The manual Black Level fix for the same unit problem (branch hl2/waterfall-manual-black) reads the same flag and merges cleanly with this one. No doc change here; that branch documents the row unit.

73, ON8ST

🤖 Generated with Claude Code

[aethersdr-agent]

1. Issue fit

Yes. No Fixes line — Refs #5678 / Refs #277 — and GOVERNANCE.md does not want an RFC for a bug fix with a clear root cause, which this is.

I verified the root cause rather than taking it from the body. RadioModel::onBackendSpectrumFrame (src/models/RadioModel.cpp:6659) re-emits the host-computed pan frame as the waterfall row ("the backend has no separate waterfall plane, so the pan row is the waterfall row", line 6706), and that lands in SpectrumWidget::updateWaterfallRow. The old baseline > 0.0f term therefore is false on every row of a negative-valued frame, so NB Blank is an inert control on HL2/ANAN/RTL-SDR. The author's "dropping the guard is not the fix" argument also holds: -147 * 1.15 = -169, so a bare ratio test blanks everything.

I also confirmed the load-bearing justification for the source-text pin: tests/tests.cmake names SpectrumWidget.cpp exactly once (in this PR's own comment), no test target compiles it, and the pattern has precedent in the file at tests.cmake:6622.

2. Scope

Everything in the diff is explained by the stated fix.

File	What it changes	Claimed?	Verdict
src/gui/WaterfallImpulseBlanker.h	New header-only decide(); tile law verbatim, dBm law added	✅	In scope
src/gui/SpectrumWidget.cpp	+1 include; inline test → one call; two identical ring writes → one	✅	In scope. Behaviour-preserving for tiles — see below
tests/tests.cmake	Registers waterfall_impulse_blanker_test	✅	In scope
tests/waterfall_impulse_blanker_test.cpp	47 checks	✅	In scope

Four files, nothing else. No CHANGELOG.md entry (correct — it must not have one). No new public/protocol/settings surface: RowKind and decide() are internal, no capability field is added, no default changes. No sibling call site left behind — m_wfBlankerThreshold has exactly one consumer and it is this one.

Deleted behaviour: the only - lines of substance are the two per-branch ring writes. They collapse correctly — decide() returns min(rowMean, baseline*1.05f) on a tile impulse and rowMean otherwise, which is byte-for-byte what the two deleted lines wrote, and the unconditional write sits in the same place the old pair did (no early return between).

Not a preference change. The old behaviour is a dead control on three families, with hardware evidence in the body. The direction (higher slider = less blanking) matches the tile law. Nothing user-visible changes on a Flex.

Socket tests: none added, removed, or modified. The new test is Qt-free and opens nothing — grep -E "QTcp|QUdp|QLocal|bind\(|listen\(|connectToHost" on it returns 0 hits.

3. Blockers

None.

4. Nits (non-blocking)

• docs/architecture/radio-capabilities-map.md:529 goes stale. It states: "panBinsAbsolute() is consumed too, and it is the second term of ONE gate." After this merges there are two consumers, and #6115's body says it "records the flag's second consumer" — so whichever lands second leaves the doc wrong either way. One line here is cheaper than coordinating the order. I looked for a canon sentence requiring a consumer to be documented and did not find one in AGENTS.md, so this is a convention, not a rule.
• The source-text pin is whitespace-collapsed exact-match. A future clang-format rewrap of those five lines turns it red with no behaviour change. You flag it yourself and offer to drop it; I would keep it — there is genuinely no behavioural seam, and the precedent exists — but it is worth a comment in the test saying it is reformat-sensitive.
• The test's Ring hardcodes kN = 32 with nothing pinning it to SpectrumWidget::WF_BLANKER_N (currently 32, SpectrumWidget.h:1992). If that constant ever moves, the ring-level tests silently stop modelling the widget.
• RowKind::Dbm slightly overclaims the unit. RadioModel.cpp:6662 calls the HL2 payload "DC-centred dBFS bins", and RTL bins are dBFS too. The law is a pure difference so it is unit-agnostic and correct either way — but the enum name asserts something the producers do not.
• Pre-existing, not yours: SpectrumOverlayMenu.cpp:2397 tooltips the slider "Higher values blank more aggressively." That is backwards for both laws. Worth naming because this PR is the first thing to state the slider's meaning in dB. Also m_wfBlankerMode ("0=Fill, 1=Interpolate") is persisted and settable but read nowhere — a dead settings key from #277.

5. What I tried to break

The new test is Qt-free, so unusually I actually ran this one rather than only reasoning about it: c++ -std=c++20 -I src on the standalone test, plus my own probe harness against the shipped header.

• Reproduced the red-before-green table independently. I applied three of your mutations myself and got your exact counts: dBm branch inert → 11 failed, finite guard dropped → 3, rejected row uncapped → 4. The clean run is 47 checks, 0 failed, with 121176 tile cases / 27040 impulses and "75 rows held" — every number in the body reproduces. The red-before-green claim is real, not decorative.
• Attacked the threshold range for a divide-by-the-margin blow-up. If threshold could reach 1.0, dbMargin is 0 and the dBm cap becomes min(rowMean, baseline) — a baseline that can only fall, i.e. a permanent freeze. It cannot: setWfBlankerThreshold clamps to [1.05, 2.0] (SpectrumWidget.cpp:4800) and the settings load clamps identically (:2511), and the slider is 5..95. Both doors are shut. No finding.
• Periodic impulses, to see whether the cap lets the baseline creep until blanking dies. At 1/2, 1/3, 1/5 and 1/10 duty with a 37 dB burst over 1000 rows: 0 missed on either row kind, baseline settling at −142.0 (dBm) and 113.7 (tile). The cap does its job.
• A slow ramp, the classic false-positive for a margin test. +0.5 dB/row for 200 rows at slider 5: dBm blanks 188/200, tile blanks 187/200. That is #277's existing slow-ramp behaviour inherited exactly, not something the dBm law introduces. At 1.15 and above, 0/200 on both. Parity, no finding.
• Hold length on a sustained step — the one stimulus you actually measured on hardware. A 36 dB step that stays is held 108 rows (4.3 s) at slider 5 and 75 rows (3.0 s) at the default, released thereafter; the tile equivalent is 86 and 55 rows. Same regime, dBm slightly longer. Maintainer-visible: the LNA step in your own evidence table will now produce a ~3 s frozen stripe where it previously passed straight through. I think that is correct — it is what a Flex already does — but it is the first thing a user will notice and it deserves to be a conscious call.
• The transition no test covers: connect/disconnect across row kinds without a restart. MainWindow.cpp:7561 sets the flag as connected && caps.panBinsAbsolute(), and the ring is never reset. Flex→HL2 leaves ~108 residue read under the dBm law: rowMean − baseline is hugely negative, never fires. HL2→Flex leaves ~−147 residue read under the tile law: baseline > 0.0f is false, never fires. Both fail open and self-heal within 32 rows (~1.3 s of under-blanking). Benign in both directions.
• Non-finite handling both ways. -inf and NaN baselines fail open on the dBm path by the explicit guard, and on the tile path incidentally (NaN > 0 is false, rowMean > +inf is false). A -inf row mean still enters the ring and disables the blanker for 32 rows — your test pins exactly that and the recovery, and I could not turn it into a permanent freeze.
• Equivalence of the tile path, beyond your grid: the collapsed ring write is the same value in the same position as the two deleted lines, with no early return between the branch and the write. I read the whole function (SpectrumWidget.cpp:8450-8690) rather than the hunk.
• Merge with #6115, which touches the same function: git merge-tree produces 0 conflict markers. Your "merges cleanly" claim holds. Worth the maintainer noticing that the two PRs split one unit bug across two new headers (WaterfallImpulseBlanker.h and WaterfallLevelMap.h) with the same extraction-plus-source-pin shape — reviewable independently, but it is one problem in two parts.
• CI is green on 63a73a35 across all five checks, including check-windows (so no MSVC surprise from the bare near() helper).
• Could not check: anything at runtime in the GUI — no build, no radio. You state this yourself for ANAN and RTL-SDR and for the fix-in-app, and I cannot close any of those gaps. The call-site wiring is reasoned from code, since nothing links SpectrumWidget.

6. Recommendation

Approve with nits. This is the most falsifiable PR body I have reviewed here — every number in it reproduced when I ran the test and the mutations myself, the tile path is provably unchanged, and the four attacks I thought would land (threshold underflow, baseline creep under periodic impulses, slow-ramp false positives, cross-radio ring residue) all failed against the code.

The one genuine judgment call is kDbPerThresholdUnit = 100, which you correctly hand to the maintainer. I would not escalate it to "needs a decision": your comparison table rules out 10*log10(t) and (t-1)*|baseline| for stated reasons, and my hold-length numbers independently show the resulting dBm margin lands in the same regime as the tile ratio rather than being arbitrary. It is one constant in one header if anyone wants it moved.

Next step: add the one-line radio-capabilities-map.md:529 correction for the second consumer, and let the maintainer rule on the sustained-step hold being the intended new behaviour on HL2. Nice work — 73.

---

_{🤖 aethersdr-agent · cost: $9.9012 · model: claude-opus-5}

[ten9876]

Issue fit

Yes, on the three families it targets, with one gap. There is no Fixes line (Refs #5678, Refs #277); a bug fix with a clear root cause needs no RFC (GOVERNANCE.md). The root cause holds. RadioModel::onBackendSpectrumFrame (src/models/RadioModel.cpp:6659-6706) reuses the backend's dB pan frame as the waterfall row. baseline > 0.0f is therefore false on every row, and NB Blank does nothing on HL2, ANAN and RTL-SDR. The fix keys the row unit on m_panBinsAbsolute. That flag picks out exactly those three families, but it does not pick out every backend whose rows are dB. See maintainer flag 1.

This review builds on @aethersdr-agent's review. I did not repeat what it reproduced. I did the three things it could not: built the branch, drove it through the bridge, and modelled cross-radio switches of the blanker ring.

Scope

File	What it changes	Claimed?	Verdict
src/gui/WaterfallImpulseBlanker.h (new)	Pure decide(). The tile law is unchanged; a dB-margin law is added	Yes	In scope
src/gui/SpectrumWidget.cpp:8640-8678	+1 include. The inline test becomes one call, and the two ring writes become one	Yes	In scope. The tile path is unchanged (see below)
tests/tests.cmake:3370-3381	Registers waterfall_impulse_blanker_test	Yes	In scope
tests/waterfall_impulse_blanker_test.cpp (new)	47 checks, no Qt, no sockets	Yes	In scope

Everything in the diff is explained by the stated fix. There is no new public, protocol or settings surface, no default change and no CHANGELOG.md entry. Commits 2891b138 (10-01) and 83c367b8 (10-03) are the fix and a comment trim, and the other two are merges from main. The test opens no socket (I read it: no QTcp/QUdp/QLocal/bind(/listen(/connectToHost).

Blockers

None.

Needs a maintainer decision

1. The predicate misses two of the five backends whose rows are dB. On an Icom (for example an IC-7300), the title's claim "NB Blank works on a dBm waterfall row" is still false. #6115 has the same gap, so between them the bug is fixed on the same three families twice. WaterfallImpulseBlanker.h:7-10 and SpectrumWidget.cpp:8646 treat panBinsAbsolute() as "this row is dB". That capability answers a different question: "do the bins hold still while the reference level moves" (RadioCapabilities.h:240-247, and the table at docs/architecture/radio-capabilities-map.md:523-527). Every backend that feeds onBackendSpectrumFrame produces dB rows:

   • Icom: IcomCivBackend.cpp:1730, spectrumFrameReady(0, floatBytes(toDbm(...)))
   • the demo: SimSignalSource.cpp:137, kFloorDbm = -120.0
   • HL2, ANAN and RTL-SDR

   Only the last three declare binsAbsolute (Hl2Backend.cpp:1767, AnanBackend.cpp:386, RtlSdrBackend.cpp:117). Icom and the demo therefore still take the tile branch with a negative baseline, and NB Blank stays dead on both. The body says so ("Icom and the demo backend leave the flag false and keep the tile test") but presents it as intended. It is the same defect. I ran the shipped header against demo-shaped rows: a ring of -120 dB rows followed by a +40 dB step, through RowKind::TileIntensity, gives 0 blanked rows in 100 at 1.05, 1.15, 1.50 and 1.95.

   #6115 (WaterfallLevelMap, manual Black Level) uses the same predicate and writes the premise into canon: "A backend with absolute bins has no waterfall plane of its own". That is true, but the converse is not, and the converse is what the code relies on. After both PRs, the Black Level slider and NB Blank on Icom and the demo stay broken in the same way.

   The partition that matches the defect is "this row came through the backend seam, not a Flex tile". Options:

   • (a) A separate declared row-unit capability, or a flag set at the place where onBackendSpectrumFrame emits the row, used by both PRs.
   • (b) Declare binsAbsolute on Icom and the demo, if their bins are in fact absolute. That has to be decided per radio, because it also opens the auto-floor gate.
   • (c) Accept the three families now and open a follow-up that names Icom and the demo.

   Whichever is chosen, #6115 and #6117 should make the same choice. I am not treating this as a blocker: nothing regresses, and the author disclosed the omission.

2. The user-visible change on HL2, ANAN and RTL-SDR is a held stripe on any sustained level step. The body's own stimulus is an LNA step. That step passed straight through before; with this PR it holds for 75 rows (about 3 s) at the default and 108 rows at 1.05. The existing review measured the same thing. A Flex already behaves this way, so this is probably correct. It is still the first thing an operator will notice, so it should be ruled on deliberately, together with kDbPerThresholdUnit = 100 (WaterfallImpulseBlanker.h:33).

Overlap with #6115

• No textual conflict. git merge-tree --write-tree 63a73a35 a54dff9c is clean (exit 0), and so is a merge of this head with current main (579efd73).
• The same root cause is fixed in two halves with matching shapes. Each PR adds a header extracted from SpectrumWidget, keys on the same m_panBinsAbsolute and adds a whitespace-collapsed source-text pin. The code is consistent: both branch on the same member, so they cannot disagree about a given radio. Point 1 above applies equally to both.
• Naming has drifted between the two. After review, #6115 renamed its parameter to rowsAreAbsoluteDb, because its body says the rows are "dBFS under a dBm label on all three families". This PR still names the same fact RowKind::Dbm (WaterfallImpulseBlanker.h:16), and its comments say "dBm row" throughout. One name for one fact across both headers would help, for example RowKind::AbsoluteDb.
• Docs. #6115 adds "It has a second consumer: the unit of a waterfall row" to radio-capabilities-map.md, and that paragraph names only the Black Level. Once both PRs merge, NB Blank is a third consumer and the doc does not mention it. The fix is one sentence in whichever PR lands second, or one PR carries the paragraph for both. The existing review's radio-capabilities-map.md:529 nit becomes this point once #6115 is in.

Nits (non-blocking)

• Correction to the existing review: "self-heals within 32 rows" is wrong for HL2→Flex, though in practice the case cannot happen. I ran the shipped decide() on a ring holding 32 rows at -147 followed by Flex tiles at 108. The baseline crosses zero at row 19, and from then on every tile row is an "impulse" capped at baseline * 1.05. The result is 169 rows blanked at 1.05 and 138 at 1.15 (about 5.5 s). Flex→HL2 blanks 0. The same arithmetic applies to main today, because main writes HL2 row means into the ring too, so this PR does not introduce it. In a real radio swap, reprojectWaterfall (SpectrumWidget.cpp:7354) resets the ring on any change of centre or bandwidth, which is why I say it cannot happen in practice. Now that the ring's unit depends on m_panBinsAbsolute, setPanBinsAbsolute (SpectrumWidget.h:362) is the natural place to make that reset unconditional:

      void setPanBinsAbsolute(bool on)
      {
          if (m_panBinsAbsolute != on)
              resetWfBlankerState();  // the ring's unit follows this flag
          m_panBinsAbsolute = on;
      }

• The Ring in the test hardcodes kN = 32 (tests/waterfall_impulse_blanker_test.cpp:37), and nothing ties it to SpectrumWidget::WF_BLANKER_N. This was already raised; I agree with it.
• The tooltip is backwards, and was before this PR. SpectrumOverlayMenu.cpp:2397 says "Higher values blank more aggressively". Under both laws a higher threshold blanks less. This PR is the first place the slider's meaning is written down, so it is a cheap moment to fix the sentence.

What I verified and what I tried to break

• Built the head (63a73a35, Linux, Qt 6 / GCC, RelWithDebInfo) and ran waterfall_impulse_blanker_test: 47 checks, 0 failed. The existing review already reproduced the mutation table, so I did not run it again.
• Bridge session on the demo (DEMO-0001). I confirmed get radio returned "model":"AetherSDR Demo","family":"sim". The session used isolated settings, offscreen, NO_TX, and my own absolute socket. I opened the Display panel and set displayWfBlankerThreshSlider=5 and displayWfBlankerBtn on, and get display read back wfBlankerEnabled:true, wfBlankerThreshold:1.05. Both values were still there after a relaunch.
• I could not grab the waterfall. The GPU renderer cannot start offscreen: get rhi returned "rendererFailed":true and grab SpectrumWidget returned "grab produced an empty image". AETHER_NO_GPU=1 failed the same way offscreen. A waterfall drawn on the CPU needs -DAETHER_GPU_SPECTRUM=OFF, which is a full rebuild I did not do. So "Flex and demo unchanged" is proven from the code, not from pixels. The demo leaves the flag false, so it runs the byte-identical tile branch, where baseline > 0.0f is false on its -120 dB rows, exactly as on main. The PR's 121 176-case equivalence grid covers the Flex tile path, and I re-read the full ring-write path (SpectrumWidget.cpp:8625-8681): the value is the same and so is the position. The demo goes through the backend seam, so this is not a Flex-wire proof either way.
• Cross-radio ring residue (nit 1), modelled on the shipped header in both directions and at four thresholds.
• Automated pass (/code-review, medium) found exactly one issue, the same Icom/demo predicate gap as maintainer flag 1. It arrived there independently. It also confirmed that HL2 bins cannot be -inf or NaN, because Hl2Spectrum.cpp adds a 1e-24 floor before the log, and that the ring-unit mixing on a capability flip already happens on main.
• Threshold floor. setWfBlankerThreshold clamps to [1.05, 2.0], so the dBm margin can never reach 0 and freeze the ring. This holds.
• CI is green on 63a73a35, but the earliest run was created at 17:04Z, and main has since moved to 579efd73 (19:25Z, 6 commits). None of those commits touch SpectrumWidget.cpp or tests.cmake, and the merge with current main is clean. So the green is very probably still accurate, but it was not produced against current main.

Recommendation: approve with nits. No blockers, so this review is a COMMENT. It is correct for the three families it claims, and the author disclosed what it leaves out. Two calls are for the maintainer: (1) the scope of the predicate, meaning fix Icom and the demo here or in a follow-up, with #6115 following the same decision; (2) the new held stripe on a sustained step, together with kDbPerThresholdUnit.

[on8st]

Bad news first: the Icom and demo gap is real and is not fixed here.

Predicate (flag 1), yours. Confirmed: IcomCivBackend.cpp:1730 and SimSignalSource.cpp:137 hand over dB rows and declare no binsAbsolute. #6115 will follow the same choice.

option	cost
(a) a row-unit capability, or a flag set where onBackendSpectrumFrame emits the row	one field or flag, both headers switch to it; reaches all five
(b) binsAbsolute on Icom and demo	also opens their auto-floor gate; unmeasured here
(c) three families now, follow-up	Icom and demo stay dead until then

(a) is the one that matches the defect; we build it on your word.

Done in efcbb06b and b75b8356, no arithmetic change:

• RowKind::AbsoluteDb, as fix(gui): manual Black Level works on a dBm row. Principle VIII. #6115.
• setPanBinsAbsolute empties the ring on a change, as proposed; the 138 held rows are now a check.
• kRingRows: the test takes it, the widget static_asserts it.
• Tooltip: "Higher values blank less."
• radio-capabilities-map.md:529 names NB Blank's use. Merges with fix(gui): manual Black Level works on a dBm row. Principle VIII. #6115 in either order.

Yours: the ~3 s hold on a sustained step, with kDbPerThresholdUnit.

The a11y scan now covers two more files: six existing warnings, none on a changed line.

73, ON8ST

[ten9876]

Maintainer ruling on #6115 and #6117: merge both as they stand. They're correct for HL2, ANAN and RTL-SDR, and the bodies now say Icom and the demo aren't covered.

For the gap, I agree with you: option (a), a row-unit signal set where onBackendSpectrumFrame emits the row, as its own follow-up PR. If it changes the demo's auto-floor, name that in the PR. Please go ahead and build it.

73, Jeremy KK7GWY

[ten9876]

We're updating this branch against main and landing it (maintainer pass); please don't push to it meanwhile. New commits go on top, with no rebase or force-push.

[ten9876]

Maintainer pass. Answers to the two reviews' points (Jeremy's review 5402519886 and aethersdr-agent's 5402028463, both body-only). The head is eeb52ce89: on8st's efcbb06b and b75b8356, plus a signed merge of current main (5fae2f863) with no conflicts.

Needs a maintainer decision

1. Predicate scope (Icom and the demo). Ruled 2026-10-03: merge as it stands, correct for HL2, ANAN and RTL-SDR. Option (a), a row-unit signal set where onBackendSpectrumFrame emits the row, comes as on8st's own follow-up PR. The PR body states that Icom (IcomCivBackend.cpp:1730) and the demo (SimSignalSource.cpp:137) are not covered. Nothing in this PR changes for that.
2. The ~3 s hold on a sustained step, and kDbPerThresholdUnit = 100. Covered by the same ruling. Unchanged.

Overlap, docs and nits: all done by on8st, verified at eeb52ce89

• Naming: RowKind::AbsoluteDb (efcbb06b), which agrees with fix(gui): manual Black Level works on a dBm row. Principle VIII. #6115's rowsAreAbsoluteDb.
• Docs: radio-capabilities-map.md now names NB Blank's use of panBinsAbsolute(). fix(gui): manual Black Level works on a dBm row. Principle VIII. #6115's paragraph is reconciled when it is updated after this one lands.
• Ring residue on a row-kind change: setPanBinsAbsolute empties the ring (b75b8356). The test checks it, and a mutation proves the check: with the reset call removed, 52 checks give 1 failure.
• kN = 32: the ring size is now kRingRows, and the widget static_asserts it (efcbb06b).
• The tooltip is now "Higher values blank less." (SpectrumOverlayMenu.cpp:2397).
• The source-text pin is kept, and its comment now says a reformat of those lines turns it red.
• Not changed, and pre-existing: m_wfBlankerMode (from Noise Blanker–Gated Waterfall Impulse Suppression #277) is persisted but never read. That is outside this fix and was not part of the ruling.

Verification on Linux/GCC (Qt 6, RelWithDebInfo) at eeb52ce89

• Full build is clean, including the new static_assert.
• waterfall_impulse_blanker_test: 52 checks, 0 failed.
• ctest excluding socket-owning tests: 100% of 480 passed. 124 tests were left out because they are socket-owning by name or by a grep of their sources.
• Static gates all exit 0: engine boundary, touchpoint manifest, capability records, command plane, test registration, CI test gate, network timeouts, qsettings, theme seed, shader dialects, qt pin and seam-probe table. The colour ratchet is +0. a11y: 6 findings, all on lines this PR doesn't touch.
• Nothing was observed in the running app (the GPU waterfall can't be grabbed offscreen).

Maintainer pass by Claude for @ten9876.

[ten9876]

Approving at eeb52ce89 (maintainer pass, pr-land).

Maintainer ruling (2026-10-03): merge as it stands. It is correct for HL2, ANAN and RTL-SDR. Icom and the demo are not covered, and the body says so. Option (a), a row-unit signal emitted from onBackendSpectrumFrame, follows as on8st's own PR. kDbPerThresholdUnit = 100 and the sustained-step hold stay as they are.

Review findings, all fixed by on8st and verified here:

• RowKind::AbsoluteDb naming: efcbb06b.
• kRingRows plus the widget static_assert: efcbb06b. The full app build compiles with it.
• Ring emptied on a row-kind change: b75b8356. A mutation proves the test catches its removal: 52 checks, 1 failed.
• Tooltip "Higher values blank less": efcbb06b.
• radio-capabilities-map.md names NB Blank's use of the flag.
• Comment on the source-text pin.

This pass: a signed merge of main 5fae2f863 (no conflicts) at eeb52ce89. All 6 threads were answered with evidence and resolved. Both body-only reviews were answered in a top-level comment.

Verified empirically (Linux/GCC, RelWithDebInfo, at eeb52ce89):

• Full build is clean.
• waterfall_impulse_blanker_test: 52 checks, 0 failed.
• ctest excluding socket-owning tests: 480 of 480 passed. 124 were excluded by name or by a source grep.
• All Static checks scripts exit 0, and the colour ratchet is +0.

Reasoned, not observed: the in-app waterfall on HL2, ANAN and RTL-SDR. There was no radio, and the GPU waterfall can't be grabbed offscreen.

Freshness: the CI run on this head was created at 2026-10-03T23:08:54Z. main's tip 5fae2f863 dates from 22:49:15Z and is in the merged tree, so there is no drift.

[ten9876]

Freshness note added after the approval. main moved to f24ac481c (#6124, 23:18:54Z) after this head's CI run was created (23:08:54Z). That commit touches MainWindow.cpp/.h, MainWindow_Session.cpp, RadioModel.cpp, tests/radiomodel_audio_mute_test.cpp and tools/check_command_plane.py, none of which this PR changes, and git merge-tree of eeb52ce89 against it is clean. On the maintainer's decision this is accepted as drift, so no further update.