profile-load hold: remaining user-intent writes still silently suppressed (census follow-up to #4201)

[Ozy311]

Summary

Umbrella follow-up from #4142/#4201. The profile-load hold suppresses every command isProfileOwnedRadioStateWrite() matches; #4201 routes the tune/zoom/band class of user-intent writes (center=, bandwidth=, band=) through the deferred pan-write queue so they replay instead of dying. A 34-writer census done for that work shows what is still silently destroyed if a user acts during the hold.

Still suppressed, by field (pan writes — mechanism ready, one-line routings + per-field dedupe semantics + hardware QA each)

rxant, wnb / wnb_level, rfgain, loopa / loopb, min_dbm / max_dbm, band_zoom / segment_zoom, fps, average, weighted_average, daxiq_channel — roughly 20 user-intent call sites still hand these to sendCommand()/sendCmd() directly, where the hold backstop drops them (now at qCDebug, by design for echo writers; these are not echo writers).

Wholesale-suppressed command families

The classifier also suppresses slice set , display panafall set , and display waterfall set wholesale during the hold — which includes user actions (e.g. a slice rfgain change) alongside the echo traffic it exists to suppress.

Known-hard site

dbmRangeChangeRequested (MainWindow_Wiring) is genuinely mixed: a user's dBm drag and the auto-floor reconcile share one wire path, so it cannot be routed as user intent without splitting the signal at the widget level first.

Note

RadioModel::setPanBandwidth() currently has no callers; if one appears it would silently drop during the hold — worth either routing or removing when touched.

Census and per-site notes live with the #4201 work; happy to break any slice of this into its own issue on request.

[aethersdr-agent]

Triage: root cause confirmed, ready for maintainer review

Verified the census against the current tree — the mechanism and the still-suppressed sites are real. Setting maintainer-review.

Root cause (confirmed)

The single choke point is the sendCommand backstop:

• RadioModel::sendCommand() — src/models/RadioModel.cpp:4848-4859 — drops any command matching isProfileOwnedRadioStateWrite() while profileLoadRadioStateWritesHeld(), now at qCDebug (no user-visible trace), and returns kProfileLoadSuppressedCommandCode with no replay.
• The classifier isProfileOwnedRadioStateWrite() — src/models/RadioModel.cpp:106-134 — returns true for display pan set <id> <k>=… (any key except the whitelisted xpixels/ypixels), and wholesale for slice set , display pan set , display panafall set , display waterfall set (lines 130-133).

So every user-intent write that is a bare sendCommand("display pan set … <field>=") / sendCmd(...) is classified as profile-owned echo and silently destroyed if the user acts inside the hold window (kProfileLoadStateWriteHoldMs). The fix pattern already exists — #4201's tune/zoom/band writes get an optimistic model update (pan->setCenterBandwidth) plus deferral/replay rather than a direct sendCommand.

Spot-checked still-suppressed sites (all direct sendCommand, all match the classifier)

• rxant — src/gui/SpectrumOverlayMenu.cpp:492
• wnb_level — src/gui/MainWindow_Wiring.cpp:3606
• rfgain — src/gui/MainWindow_Wiring.cpp:3614
• daxiq_channel — src/gui/MainWindow_Wiring.cpp:2757
• band-stack replay path also fires these: slice set … rxant= (MainWindow.cpp:4000), wnb_level= (MainWindow.cpp:4038)

The dead-code note checks out too: RadioModel::setPanBandwidth() (RadioModel.cpp:2738) has zero callers — only its decl (RadioModel.h:444) and def. It would silently drop during the hold if a caller ever appears.

Proposed fix — slice this umbrella into three PRs

The reporter already offered to split; recommend doing so along the natural risk boundaries:

1. Pan-write field routings (~20 sites, mechanical). Route each display pan set … <field>= user-intent site through the same deferred/replay path fix(spectrum): defer, never drop, profile-owned pan writes (#4142) #4201 established, with per-field dedupe (last-write-wins per (panId, field)) and an optimistic model update where a canonical value exists. Each field needs a hardware QA pass (rxant, wnb/wnb_level, rfgain, loopa/loopb, min_dbm/max_dbm, band_zoom/segment_zoom, fps, average, weighted_average, daxiq_channel). Good incremental PR, low structural risk.

2. Wholesale command-family narrowing. slice set , display panafall set , display waterfall set currently suppress user actions (e.g. a slice rfgain) alongside the echo they exist to filter. Needs the classifier tightened to distinguish echo vs user intent (likely a caller-side "this is user intent" tag rather than string-sniffing at the backstop), so higher design risk — keep separate.

3. dbmRangeChangeRequested (known-hard). src/gui/MainWindow_Wiring.cpp:2440-2489 — user dBm drag and the auto-floor reconcile share one wire (sw->pendingAutoNoiseFloorDbmRange() already partly disambiguates at 2451). Cannot be routed as pure user intent without splitting the signal at the widget level first. Own issue.

4. setPanBandwidth() dead code — route or delete when next touched; trivial, fold into (1).

Handoff summary

Root cause is a single backstop + over-broad classifier; the deferral/replay + optimistic-model fix pattern is already proven by #4201, so this is extension work, not new design (except family #2 and site #3). Recommend authorizing PR (1) first as the bulk, low-risk win, and filing (2)/(3) as separate tracked issues. No safety/TX-power implication — bounded to display/gain state during the profile-load settle window.

---

_{🤖 aethersdr-agent · cost: $1.5838 · model: claude-opus-4-8}

---

_{📍 Citation health check (Foundry Constitution Principle I): 7 resolved, 4 unresolved against /Users/aetherclaude/workspace/AetherSDR at origin/main.}

Unresolved citations (4)

Citation	Reason
MainWindow.cpp:4000	file not found in workspace
MainWindow.cpp:4038	file not found in workspace
RadioModel.cpp:2738	file not found in workspace
RadioModel.h:444	file not found in workspace