[RFC] Complete RTL-SDR receive DSP with multiple slices, independent zoom and stereo WFM #5468

Current rebased stack — 2026-10-06: main 5f09371e → #5973 (1398f3b9) → #5974 (ebb35cfd) → #6082 (c667b41c) → #6083 (ca3a36bd). Separate Digital draft #6002 follows at 5ae6b202; analog does not depend on it. Upstream window chrome and release debug-symbol changes are preserved.

Fresh rebase validation on analog ca3a36bd: native Apple Silicon / Qt 6.12 / RTL required / Digital OFF application build succeeds. 51/52 selected CTests pass in 226.55 seconds: all 47 RTL regressions and four additional UI checks pass. The new upstream unified_title_bar_test fails its Alt+F/menu-rounding assertions under Mac offscreen rendering; an independent rebuild and run on clean main 5f09371e reproduces the identical two failures. This is an inherited test failure, not a green full-suite claim. Static engine/routing/capability, generated manifests, registration/frozen-gate, theme and build-options checks pass; package-helper tests pass 9/9 and required-RTL configuration checks pass 16 combinations. The Digital increment and vendor trees are unchanged by replay; decoder-ON tests and release packaging were not rerun for these heads. Hosted checks are being refreshed by the push.

Published analog stack reconciled: #5924 has merged; the remaining order is main 035a8583 → #5973 (45028ceb) → #5974 (b8aba26e) → #6082 (1e0276fc) → #6083 (dd6be71e). Only the individual increments were replayed, preserving main’s final SQL/FM rules, typed receive routing and current architecture ratchets. The DSP foundation’s Linux/macOS/Windows CI passed. Focused Mac checkpoint tests passed for DSP (1/1), WFM/audio/filter integration (4/4) and PPM (3/3). The combined Mac application builds and 47/47 selected regressions pass. The Linux full suite passed with 659 passed, five skipped, zero failures; AppImage x86/ARM, macOS DMG Apple Silicon/Intel, and Windows packaging all passed. Runtime/package evidence is pinned to 50e2659e; final dd6be71e changes only BUILD-OPTIONS.md and its static gate passes. Artifacts retain their tested revision, and no new live RF or capacity qualification is claimed.

Separate Digital draft: #6002 is restacked at ac62e478 under RFC #6079 and remains draft. Nobara/GCC/Qt 6.12 focused checks pass with the decoder OFF (2/2) and ON (4/4). Tested d2c61b01 differs from final Digital only in BUILD-OPTIONS.md. These are focused offline checks, not full Digital release or live-audio qualification. Analog does not depend on Digital approval. Prior integration history remains on codex/rtl-digital-evaluation-preserved-20261001.

Research date: 2026-09-06.
Related work: RTL backend issue #4797, merged experimental backend PR #4862, the existing Hermes-Lite/ANAN WDSP receive paths, and the aetherd backend boundary.
Evidence baseline: AetherSDR commit e60c3bea2ac8e5b762be9a3c6f52aa1bab8fb9ad. This is the tested 26.9.2-main portable-build source, not a claim that the tests ran on the release tag itself.

This proposal requires an RFC under GOVERNANCE.md because it changes the shared slice lifecycle, audio pipeline and receiver controls. Architecture approval was recorded on 2026-09-07 UTC. At the original proposal stage, completed work was source research, isolated feasibility probes and implementation planning. Current application checkpoints and their limited qualification are recorded above; the original acceptance requirements below remain the baseline.

Problem

The experimental RTL backend can receive WFM, but several advertised receive modes and controls are incomplete. In the reported case, selecting FM at 144.390 MHz after WFM left a 200 kHz passband displayed, and the panadapter could not zoom closely enough to inspect the channel. Source inspection and generated-IQ probes established that filter edges are stored but not applied by the current RTL demodulator, USB/LSB do not isolate a sideband, and CW lacks the required pitch offset. FM/AM do have processing branches; it would be inaccurate to describe the backend as WFM-only.

The required result is an RTL receiver that can place multiple independently demodulated RX slices across one captured IQ window, with working modes and filters, useful spectrum zoom, and broadcast-FM stereo. A user must be able to monitor several channels simultaneously without moving one receiver changing another's station or zooming the display reducing its available samples.

One RTL dongle provides one tunable RF capture. Software receivers may be separated as far as their complete passbands fit within that capture's usable bandwidth. They share the tuner center, hardware sample rate, RF gain, oscillator correction and front-end overload behavior. They do not provide reception across unrelated RF bands outside that window.

This builds on #4797 and the broader auxiliary-receiver request #4029. The decision here is how to complete the existing RTL backend with shared-window multi-RX, independent spectrum zoom and broadcast stereo.

Proposal

Feasibility evidence. A real RTL capture reached the existing WDSP stereo decoder offline. Generated stereo measured 42.30/48.43 dB separation; generated shared-IQ tests ran four chains at 7.20× real time with two tuning positions 2.1 MHz apart. These are component experiments; integrated multi-RX, live narrow-FM intelligibility and native Windows stereo remain unverified. Full measurements and limitations are retained in Appendix A.

Reuse the existing WdspChannel and r8brain infrastructure. Follow Hermes' established per-receiver ownership, stable IDs, mode/filter ordering, CW translation, DSP lifecycle and audio-mixing behavior. Add the RTL-specific shared-IQ channel extraction and shared-tuner rules described below.

Multiple RX belongs in the initial data model and processing topology. Do not replace the current demodulator with another hard-coded single-slice implementation and defer multi-RX to a later redesign.

The proposed initial acceptance target is four simultaneous RX slices on Nobara-Dell, with one visible panadapter and independently movable slice markers. The final advertised receiver limit is set per architecture from component probes and integrated performance/resource tests, including native Linux aarch64 on representative Raspberry Pi-class hardware. The x86-64 four-slice target is not an ARM default; macOS arm64 evidence does not establish Raspberry Pi performance. The eight-chain experiment below is feasibility evidence, not approval to advertise eight fully supported receivers.

The original approved analog proposal adds no third-party DSP dependency: WDSP 2.00 and r8brain are already in AetherSDR. Optional Digital FM and its additional dependencies are outside this RFC and tracked separately in #6079. RTL remains receive-only.

Capture-to-DSP boundary. The reusable software-receiver chain consumes normalized complex float IQ with explicit achieved sample rate, RF center, usable frequency interval, stable capture identity and discontinuity/reconfiguration information. Device sample conversion, hardware controls and callback lifetime remain in the RTL adapter. The chain must accept variable input block lengths and rates; the tested 2.4 MS/s setting and four-slice target are validation points, not constants in a shared DSP interface. Apply shared-tuner constraints per capture endpoint, rather than assuming every endpoint in a backend must share one center. This RFC still implements one RTL capture per session. Keep channel extraction replaceable if wider-rate measurements later justify staged decimation or a channelizer.

Related architecture issue #4955 proposes a backend-neutral IQ provider for external consumers, including canonical orientation, endpoint identity and bounded delivery. Align the internal capture contract with those concepts; this RFC does not implement its TCI subscription/export work. #3894 addresses standalone receive-session startup, a separate concern from completing this existing backend's DSP.

1. Separate capture, viewport and slice state

Maintain three distinct concepts:

• Capture: actual hardware center and sample rate, with a documented usable frequency interval and margins.
• Viewport: displayed spectrum center/span, entirely within the captured interval; changing it does not implicitly change the USB sample rate.
• Slices: stable IDs and independent frequency, mode, filter, AGC, squelch, audio level/mute/balance and DSP state.

For slice i, let its required RF interval be [f_i + low_i, f_i + high_i]. The full interval, including required transition/guard margins, must fit within the usable capture interval. WFM requires its full channel width; checking only its center frequency is insufficient. CW's operator-facing carrier interval and internal BFO-shifted DSP interval must be translated consistently before validation.

Proposed tuning behavior:

1. Keep the hardware center fixed for slice moves that fit the existing capture.
2. If a requested move requires recentering, find a legal center that preserves every active slice's requested absolute RF interval. Retune once only if the full set fits and the hardware can accept it.
3. If the set cannot fit, reject the requested move with a useful explanation; preserve the other slices and the prior valid state. Do not silently evict, retune or widen another receiver.
4. Apply the same whole-set validation to explicit hardware sample-rate changes. Publish actual successful hardware readback; roll back a failed operation coherently.
5. Keep spectrum-only zoom independent of capture and of receivers outside the currently visible subsection. Recenter/zoom UI details and refusal presentation require maintainer agreement.

The actual usable margin must be measured and specified. A 2.4 MS/s setting is not a promise of 2.4 MHz of uniformly clean usable bandwidth. RTL's current 0.45 × sampleRate center-only guard is not sufficient as the multi-slice policy.

1.1 Slice persistence and identity

RtlSliceSettings is the single owner of feature document RtlSlices, with kSchemaVersion = 1, following Hl2FilterBoardSettings. It receives the scope through RadioModel::settingsScope(). Consumers read the effective exact → family-wide → empty document; writers read-modify-write featureExact, preserve unrelated fields, refuse a newer schema, check the atomic write result and issue a qWarning on refusal. An existing empty document is a settled document, not permission to import old state again.

Use the reported serial as radioId only when it is non-empty and is not the synthesized USB-index fallback; otherwise use the RTL family-wide row (empty radioId). A connection locator such as rtl:<index> remains useful for this scan/connection, but is not persistent device identity. Preserve that provenance through discovery, connection metadata and the backend. Changing only RtlSdrBackend::m_serial is insufficient: RadioModel::settingsScope() currently reads m_lastInfo.serial, and restore is handed to the backend before connect. A genuine numeric serial must not be mistaken for the index merely because the strings match. Blank/index-fallback identities deliberately share the family-wide row. Duplicate nonempty factory serials can still collide under the stated rule; this revision does not claim per-dongle isolation or introduce another identity scheme.

The document shape is:

{
  "captureCenterHz": 100000000,
  "sampleRateHz": 2400000,
  "slices": {
    "0": {
      "freq": 100000000,
      "mode": "FM",
      "filterLow": -8000,
      "filterHigh": 8000,
      "agcMode": "med",
      "agcThreshold": 65,
      "squelch": { "enabled": false, "level": 20 },
      "audioGain": 50,
      "audioMute": false,
      "audioPan": 50
    }
  }
}

This is a schema example, not a new set of defaults. Frequencies and filter offsets in this document are in Hz; conversion to existing MHz model APIs occurs at an explicit boundary. Mode and AGC mode use normalized model strings; gain/pan are percentages and other values use the validated slice-model units. squelch contains both enablement and level, so restart does not lose half of that control. The WFM control PR adds the optional selected wfmDeemphasisUs to this same slice object, validates it against the supported capability values, and documents its reviewed runtime default; no new flat setting is introduced. Pilot/stereo detection is live state and is never restored from disk. slices is keyed by stable UI slot ID, not array position or WDSP channel number. Removing a middle slice never renumbers surviving slices. Reuse a free slot only after retirement, with a separate generation token to reject delayed work for its previous occupant.

Restore parses and validates the entire document before applying any slice. Against the current successfully established capture and its actual readback, restore the subset whose complete passbands fit, within the measured architecture limit; report each rejected slice with qWarning. Use ascending stable IDs to make capacity reduction deterministic. Never silently move the capture, reduce a slice's passband, or move a sibling to make saved slices fit. Saved capture center/rate record the saved set's context; they do not authorize a post-connect recenter to accommodate an otherwise refused restore. If no saved slice is acceptable, retain the session's valid initial state and report the refusal.

The existing RTL OperatingState currently owns tuning, passband and span. Transfer those overlapping fields to RtlSlices with a one-time, scope-safe claim migration; retain non-overlapping RF-gain ownership. Do not leave two writers or two restore paths for the same RTL slice/capture state. Do not import an index-keyed legacy row into an asserted per-device identity. A refused migration write retries later and never marks the migration complete. Persist only accepted state after successful preparation/hardware readback, not pending or rejected requests.

2. Build software receivers from one IQ stream

Conceptual receive path:

flowchart LR
    U[One RTL USB IQ stream] --> S[Shared capture block]
    S --> V[Viewport shift and filtered decimation]
    V --> F[FFT and waterfall]
    S --> A[RX A: shift, filter and resample]
    S --> B[RX B: shift, filter and resample]
    S --> N[Other bounded RX slices]
    A --> DA[Independent WDSP channel A]
    B --> DB[Independent WDSP channel B]
    N --> DN[Independent WDSP channels]
    DA --> PA[Per-slice audio]
    DB --> PB[Per-slice audio]
    DN --> PN[Per-slice audio]
    PA --> M[Aligned speaker mix]
    PB --> M
    PN --> M

Use a phase-continuous frequency shifter and identical, independent I/Q resamplers for each slice. AetherSDR's Resampler::processStereoToStereo() averages its inputs and is unsuitable for either complex IQ or preserving broadcast stereo; use the existing paired-instance pattern.

Proposed mode-rate policy follows the WDSP 2.00 manual: 192 kHz DSP for FM-family modes and 48 kHz for other analog modes. WFM specifically requires 192 kHz DSP and at least 192 kHz input. The tested path converts RTL IQ to 192 kHz before WFM; the final anti-alias specification must preserve the required channel. Some RTL rates do not satisfy WdspChannel's integral-rate/block contract directly, which is why the resampling stage is required.

Recover USB/LSB, AM/SAM, FM/narrow FM, CW/CWR and WFM through appropriate existing DSP facilities and explicit mode mappings. Preserve user-facing vocabulary where possible; resolve the current FMN/NFM and CWR/CWL differences deliberately. Apply mode changes before reapplying the corresponding passband, as Hermes does. Adopt a mode-appropriate filter on a genuine mode transition and preserve manual edits while that mode remains selected. Final preset widths and FM deviation values need reviewer agreement and test coverage.

CW requires the established BFO/filter translation; selecting a WDSP CW enum alone does not create the expected pitch. SAM must be validated for synchronous tracking, not just recovery of a centered AM carrier. FM and narrow FM must have meaningful, documented differences rather than identical branches.

Verify RTL IQ orientation independently. The generated WDSP probe selected negative mathematical complex frequency in USB; Hermes and ANAN document their own orientation handling. Tuning sign, spectrum direction, sideband selection and CW pitch must agree for RTL rather than inheriting a hardware-specific conjugation from another backend.

3. Reuse Hermes' lifecycle without copying its transport assumptions

Hermes hardware sends separately tuned DDC streams, and adding a receiver changes its EP6 wire layout. RTL distributes one unchanged capture to multiple software extraction stages. Opening another RTL slice should not require restarting USB acquisition merely to change a receiver count.

Hermes currently creates a receiver through Add Panadapter, with one receiver per pan. The requested RTL interaction is several slices on one pan. RadioModel::addSlice()/addSliceOnPan() still construct Flex commands, and the neutral backend interface has pan lifecycle methods but lacks an equivalent slice-create/remove path. Add a narrow backend-neutral slice lifecycle seam that preserves existing Flex behavior and reports normalized slice events. Raising maxSlices alone is insufficient.

Do not copy Hermes' blocking Qt event-loop synchronization into librtlsdr's asynchronous read callback. The implementation design must specify off-callback DSP preparation, safe active-list publication, retirement acknowledgment and destruction, bounded pending work, and cancellation of stale connection/reconfiguration results. Keep FFTW planning and allocating control operations away from live sample processing. Reuse the existing RTL worker where practical; an extra Qt thread per slice is not assumed. WDSP already owns internal workers.

Cold-start strategy: use bounded background preparation on the existing Qt worker-pool infrastructure, with one executing RTL preparation job and one coalesced desired configuration across active and retiring connection generations. Retain the preparation controller until its work finishes, so repeated reconnects cannot enqueue unbounded obsolete FFTW jobs. Keep WDSP's existing global setup/planner serialization. Import production FFTW wisdom and prepare the requested receiver configurations off both the GUI thread and USB callback; cache successful production plans using the existing atomic wisdom writer. Publish only fully prepared chains at a sample-block boundary, acknowledge retirement, and destroy retired chains off the sample path. Coalesce pending control changes and reject stale connection/slice generations. Canceling a request invalidates its result; it does not assume a running FFTW call can be interrupted. Keep its owner alive until completion.

A cold first connection can report that receivers are being prepared while the UI remains responsive and cancellable. Adding a receiver must not stall existing reception for planning or restart USB merely because the receiver count changes. Plan construction, mode/filter reconfiguration, buffer allocation and destruction all remain off the callback. The experimental AETHER_WDSP_FFTW_TIMELIMIT=0.001 is not the shipping strategy. For cold/warm validation, use an explicit absolute task-owned AETHER_WDSP_WISDOM_DIR that persists across warm runs and verify the resolved path. PR #5417 has merged; follow its corrected docs/HERMES.md recipe rather than redirecting HOME and assuming the WDSP cache was isolated.

Any common helper extraction should be small, with behavior-preserving tests for Hermes. An RTL feature should not require a broad rewrite of working Hermes or ANAN backends.

4. Provide broadcast stereo and preserve independent audio routes

Use WDSP's existing WBFM stereo decoder, internal squelch and deemphasis. Expose stereo detection and regional deemphasis through AetherSDR's approved WDSP C boundary and wrapper, not by including vendor headers above the DSP boundary. Provide truthful stereo/mono indication and tested mono fallback after pilot loss. The placement and styling of new controls/indicators require maintainer review.

Approved audio target: 48 kHz stereo for WFM, preserving the normal broadcast audio band through 15 kHz. Current backend PCM is 24 kHz stereo: it can preserve left/right separation, as tested, but cannot carry audio through 15 kHz. Emitting 48 kHz bytes through an API that assumes 24 kHz would misinterpret timing.

Propose a rate-aware backend audio path with explicit frame metadata for sample rate and channel layout, preserving existing 24 kHz producers. All affected buffering, playback resampling, recording, DSP taps and supported downstream routes must interpret that metadata. Convert at explicit boundaries; do not globally change AudioEngine::DEFAULT_SAMPLE_RATE or silently bypass the normal engine. A mixed narrow-mode/WFM speaker mix must use one agreed output rate while retaining independent per-slice streams.

48 kHz is approved; do not ship a new 24 kHz interim milestone. Keep the existing WFM implementation usable during development until its 48 kHz replacement passes comparison tests.

The rate-aware PR must explicitly inventory all four AudioEngine.cpp sites Jeremy identified at baseline e60c3bea: the kTargetHz = 24000 at line 1774, the main RX L/R resamplers at 4328/4331, and the RADE resampler at 9444. m_rxOutputRate is the negotiated output-device rate, not the incoming backend producer rate. The main RX L/R resamplers must use the validated producer rate. The other two sites have different contracts: line 1774 adapts CW sidetone into the fixed-rate CW decoder, and line 9444 adapts fixed-rate decoded RADE speech. Keep those explicit fixed-rate boundaries and regression-test them; do not blindly replace all four literals with one variable.

Inventory queue/drain timing, client RX DSP preparation, pre/post-DSP taps, per-slice streams, TCI/DAX conversion, QSO recording/playback and clock/decoder consumers as well. Carry sample rate, channel layout, sample-frame count and session/format generation consistently across queued delivery. Preserve L/R with paired resamplers at stereo conversion boundaries. Existing fixed-rate consumers receive explicit conversion; 48 kHz must survive through the intended WFM speaker/recording/streaming paths, not be reduced to 24 kHz and relabeled later. Existing Flex 24 kHz and Hermes/ANAN behavior receive regression coverage in the same affected PRs.

Normalize output levels before the final limiter. Raw WDSP output in the recorded broadcast probe peaked at about 3.97, with 30.3% of samples above unity. Feeding that directly into RTL's current clamp would distort it. Gain, headroom, clipping behavior and mode-transition loudness need explicit validation. Preserve unmixed per-slice audio for supported decoder/streaming routes; adding a RadioReference upload service or a new network streaming protocol is outside this RFC.

5. Support narrow spectrum views independently

Provide filtered software decimation for the selected viewport while retaining the wide capture for all receivers. A proposed initial close-view target is 24 kHz or narrower, subject to reviewed UI limits and measured spectrum behavior.

For context, a 2048-bin FFT has about 1172 Hz grid spacing over 2.4 MHz and 11.7 Hz spacing over 24 kHz. These figures describe bin spacing, not effective window resolution. Merely stretching or cropping the current coarse FFT does not add frequency resolution. Preserve calibration, axis orientation and frame pacing through viewport changes.

6. Capability and accessibility contract

Proposed additions to RadioCapabilities, with conservative defaults:

Set every new field explicitly in all six backend capability constructors, add a row to docs/architecture/radio-capabilities-map.md, update the header's ADDING-A-FIELD obligations, and add behavioral gating tests including RTL → other-family → RTL swaps and disconnect. Inspect pending backend additions when the field lands. Flex advertises the lifecycle capability with its existing semantics; do not infer other capabilities from family names or maxSlices. RTL only raises each feature flag when the corresponding implementation is complete. The advertised maxSlices comes from the measured architecture profile, not from a new universal four/eight constant.

Capability is support, not current reception status. Actual pilot/stereo detection is live per-slice normalized state, cleared on leaving WFM, loss of pilot where appropriate, removal and family swap. Selected deemphasis is persistent slice configuration: preserve it across signal loss and mode round trips, while marking it inactive and hiding its control outside WFM. Removal/family swap withdraws the old slice's live controls rather than transferring its selection to an unrelated slice. UI consumes model/capability state, never a new vendor include. Placement and final preset/deemphasis defaults remain concrete PR-review decisions within the approved direction.

Read docs/a11y.md and docs/style/theme-style-guide.md before UI work. New stereo/pilot indication, deemphasis selection, per-slice audio controls and zoom must have accessible names; every new interactive control must expose settled value changes with QAccessibleValueChangeEvent. Provide keyboard operation and appropriate accessible state/name events for indicators, plus QAccessibleInterface for custom painting where needed. Do not add interactive QLabels. Verify these behaviors even though the existing accessibility lint is warning-only: accessibility is a release gate for this RFC.

Cross-platform impact

WDSP and r8brain already build on all three supported desktop platforms, and the stereo configuration was exercised on Linux and macOS. Windows still needs a native configuration probe and integrated validation. Native Linux aarch64/Raspberry Pi multi-chain measurements are also outstanding. Run at the intended 192 kHz FM-family DSP / 48 kHz stereo output, identify CPU/board/cooling and build settings, and measure separately from macOS arm64. If representative ARM hardware is unavailable, report that gap to the maintainer and obtain a tester; do not infer a supported limit from the desktop x86-64 result. The original analog scope proposes no new license-bearing dependency; optional Digital dependencies are outside this RFC in #6079. Any vendor-source change must follow the existing vendor patch/provenance process.

The RTL dependency remains optional and the receive backend must compile out cleanly. Linux packaging must address USB driver ownership: the tested auto-detach library setting is a candidate, with confirmed reattachment on clean close. macOS and Windows bundles still require their established RTL/USB and Qt runtime dependencies. Do not change drivers or system configuration as an automatic installation side effect without an approved platform-specific plan.

A rate-aware audio seam affects shared engine consumers even though only RTL initially needs this WFM configuration. Preserve old producer behavior and keep engine implementation below the GUI boundary. New controls use normal capability/model state and the theme guide. The proposal does not alter Flex protocol commands, RF transmit behavior, or existing radio default modes/rates.

Alternatives considered

Implementation scope

The work touches src/core/backends/rtl/, src/core/dsp/WdspChannel and its approved WDSP C boundary, IRadioBackend, RadioModel/slice capability and lifecycle routing, AudioEngine and rate-sensitive consumers, spectrum processing, and the GUI connections for slice creation and controls. Tests belong in tests/ with registration in tests/tests.cmake; packaging changes are limited to the existing optional RTL dependency and its runtime requirements.

New dependencies: none proposed. The largest shared risks are audio-rate interpretation and slice lifecycle/tuning correctness. Existing Flex audio and Hermes/ANAN receive behavior need regression coverage. The following criteria and staged scope incorporate the maintainer's conditions; they are not evidence that the gates have passed.

Acceptance criteria and remaining validation

The following are required release gates, not claims of completed tests.

Use socket-free injected IQ and state/lifecycle tests for deterministic DSP and policy claims. Do not add a synthetic radio peer to the default test graph. Use the automation bridge for GUI/model convergence and receive-only radiocert for positive real-hardware receive liveness. Its current receive routine assumes slice 0, a WWV reference and specific modes; add capability-aware target/reference selection and avoid probe retunes that exceed the accepted shared window. Record unsupported or unobserved checks explicitly: radiocert is a diagnostic, not a pass/fail certificate, and a quiet RF interval cannot prove refusal, isolation or packet decoding. Apply sanitizer/lifetime coverage to receiver churn. Existing tests to study include hl2_receivers_test, hl2_receiver_churn_test and hl2_shift_test; their source was inspected, not rerun as part of this feasibility work.

Approved decisions and implementation sequence

Jeremy approved all six direction decisions in the linked comment, with the conditions incorporated above. No second architecture approval round is requested:

1. Multiple independent RTL slices on one capture/pan, with four RXs as the initial Nobara target and bounded per-architecture limits based on integrated evidence.
2. Reuse of WdspChannel, paired r8brain conversion and Hermes lifecycle/audio patterns, with minimal shared-helper extraction reviewed by Pat and the maintainer.
3. A neutral slice lifecycle seam and the proposed full-passband shared-tuner acceptance/refusal behavior.
4. Mode-specific DSP rates/presets and IQ/CW conventions, with the manual's 192 kHz FM guidance explicitly considered rather than blindly inheriting Hermes' current fixed 48 kHz configuration.
5. A rate-aware 48 kHz WFM audio path, including mixed-rate per-slice routing/mixing. Skip the 24 kHz interim.
6. Independent spectrum zoom, pilot indication/deemphasis UI, and the proposed validation gates.

Retain the approved four-stage sequence, split into one feature and its tests per PR:

• Foundation: separate PRs for the neutral slice lifecycle (with Flex regression evidence), shared-window policy, identity/scoped persistence ownership, and testable capture/processing lifecycle. Do not combine neutral lifecycle and tuning policy into one PR.
• First usable multi-RX build: working FM/channel filtering, independent audio controls/routes and bounded multi-RX integration, targeting four on Nobara and retaining the existing WFM path. The 48 kHz metadata/consumer foundations land before enabling a new producer; there is no replacement 24 kHz interim.
• Mode/zoom completion: remaining analog modes, CW/SAM verification, independent high-resolution spectrum views and full persisted-state round trips through the foundation owner.
• Stereo completion: verified 48 kHz end-to-end audio, stereo telemetry/fallback/deemphasis and output-level validation; accessibility, bridge/radiocert, all-platform packaging, architecture-specific CPU/headroom and soak gates.

Scope ends at a reliable, receive-only multi-slice RTL backend using one dongle/capture per session. Multiple dongle aggregation, coherent diversity, new transmit capabilities, new streaming services and unrelated backend rewrites are not proposed. Keep the existing working WFM path available during development until the replacement passes comparison tests.

The evidence supports feasibility; it does not establish a finished implementation or a delivery-time estimate. No fundamental hardware, algorithm, dependency or Nobara CPU blocker has been demonstrated. The largest remaining risks are shared audio-rate integration and lifecycle/tuning correctness under active multi-RX use.

Appendix A — Measurements supporting feasibility

These are completed experiments using unchanged application DSP sources in standalone probes. There is no integrated RTL-to-WDSP multi-slice application build yet.

Test environment and method

• Application baseline: e60c3bea2ac8e5b762be9a3c6f52aa1bab8fb9ad.
• WDSP vendor baseline: 2.00, upstream revision 584e8aca5ba1c4c6bc66fc0cc164ce567c8ba1e3.
• Hardware host: Nobara-Dell, Nobara Linux 44, x86_64. One Realtek RTL2838UHIDIR with R820T tuner was assigned exclusively for receive development.
• Cross-platform component check: macOS arm64, same application source and generated stereo vector. No new Windows stereo/multi-RX probe was run.
• wdsp_probe.cpp compiles unchanged WdspChannel.cpp against the app's existing WDSP archive. resample_probe.cpp compiles unchanged Resampler.cpp. multirx_probe.cpp combines both with standalone per-RX software mixing.
• Tests set AETHER_WDSP_FFTW_TIMELIMIT=0.001 and an isolated wisdom directory. This bounds FFTW planning and prevents writes of rushed plans to the user's normal cache. It is not a shipping-build configuration, and the recorded setup times do not establish production cold-start latency.
• USB capture was sequential and bounded. Recorded RF was processed offline; no transmitter was operated and no production stream was interrupted.

RF acquisition and capture results

Linux initially refused to claim the dongle because dvb_usb_rtl28xxu was active. A task-local RTL library rebuilt with DETACH_KERNEL_DRIVER=ON successfully detached the driver, opened the dongle and reattached it on close. No system blacklist, udev rule, EEPROM or bias-tee setting was changed. The existing portable package was not replaced during research.

Generated stereo, mode and filter results

The final stereo vector uses a 19 kHz pilot and a correctly phased 38 kHz suppressed subcarrier according to ITU-R BS.450-4 §2.2.2.5. It carries a 1 kHz left tone and a 2 kHz right tone. An earlier exploratory vector used the wrong quadrature relationship; its initial lack of separation was a test-vector error, and those results are superseded.

The allocation guard measures WDSP-tracked allocations, not all allocations in a future application pipeline. Centered-carrier AM/SAM similarity is not evidence that synchronous AM is fully validated. CW has not yet been demonstrated by this probe.

Processing times

Ratios use unrounded internal timings and may differ slightly from division of the displayed rounded durations. These tests show component headroom, not full application CPU utilization or latency. The generated tests include signal-generation/measurement work; the recorded-IQ test includes file I/O. They exclude USB, full GUI rendering and speaker mixing.

Shared-IQ multiple-RX results

One generated 2.4 MS/s stream contains two equal-strength FM signals with different audio tones. Every receiver sees the same input block, shifts its selected channel to baseband, uses paired r8brain conversion and runs an independent WDSP FM channel at 192 kHz DSP / 24 kHz output.

Each chain emitted 95,744 output samples with zero failed WDSP blocks. Desired-tone versus other-channel-tone ratios were approximately 81.17 dB and 81.08 dB at the two tuning positions. The output count reflects finite buffering in the short probe; it is not a long-duration clock-drift result.

Four/eight chains duplicate the two tuning positions for load testing. These runs do not prove four/eight distinct physical signals, mixed-mode reception, live retuning, receiver churn or an integrated maximum receiver count. The measured section includes per-chain frequency mixing, resampling and demodulation; it excludes input generation, setup, USB, FFT/display work, speaker mixing and GUI activity. The current retained probe source uses the final ±1050 kHz case; reproducing the earlier table rows requires setting its two offsets back to ±800 kHz.

RTL-rate conversion coverage

The existing paired resampler was exercised at every advertised rate, converting to 192/48/24 kHz. The two identical input streams produced byte-identical outputs in every run; output duration matched input duration to displayed precision. These constant-input runs measure throughput and paired behavior, not all-rate RF rejection.

Appendix B — Source and evidence index

Application source links below are pinned to the tested commit:

• RTL backend and capability/tuning state, RTL demodulator, experimental design/limitations.
• Hermes receiver lifecycle, mode/filter/CW handling and mixer, receiver IDs, Hermes RX DSP, Metis per-DDC IQ demultiplexing.
• Shared WDSP wrapper, paired-resampler building block, audio engine and 24 kHz assumptions.
• Current slice-create routing, neutral backend interface.
• Bundled WDSP stereo decoder, WDSP vendor provenance.
• WDSP Guide 2.00, §§5.1 and 5.3.10, ITU-R BS.450-4.

Companion evidence bundle contains FEASIBILITY.md, the three standalone C++ probes and CMake definition, analyze_captures.py, final result logs, USB/capture logs, RF survey, and the extracted upstream guide. Numeric results are included above so this draft remains understandable without private host access. Raw RF and intermediate audio are retained on the development host; no public attachment or upload is implied by this draft. Hashes and reproduction notes accompany the local bundle.

73, Ozy K6OZY · GPT-6 Astra-Ultra

Operator amendment: eight-receiver Nobara evaluation and SQL correction (2026-09-28)

The operator now requests eight simultaneous analog receivers in the isolated Nobara evaluation, replacing the earlier four-receiver evaluation target. Their approximately 3% laptop CPU observation with four is context, not a measured eight-receiver capacity or release qualification. This amendment is recorded for maintainer review; it does not claim maintainer approval, change the qualified/default capacity, authorize multiple digital decoders, or expand Mac/Windows/public distribution.

The immediate defect report is high manual SQL values needed to silence noise and non-working Auto SQL. Preserve the operator's before-state: A 460.300 MHz FM-N SQL 54; B 460.275 MHz FM SQL 51; C 460.1754 MHz FM SQL 51; filters ±8 kHz, all unmuted/volume 100, B selected, saved Auto off. The identified private instance was captured read-only, its database backed up consistently, and then closed cleanly with operator authorization. RF gain was 23 dB, capture 460.9 MHz / 2.4 MS/s; these are observed settings, not calibrated RF measurements.

Source diagnosis found that the 65536-bin display estimator drives a 2048-bin audio detector despite their different noise-bin power, and that a trimmed logarithmic noise average is biased low for a peak gate. The correction under focused validation moves RTL Auto measurement to each receiver's existing acquisition gate using its detector spectrum, a local noise estimate excluding the desired passband, and independent Auto/margin state. It uses typed model/backend control and accepted per-slice persistence. Manual -120 + 1.2 × level mapping is retained pending real station comparison; no guessed offset is applied. WFM keeps its separate decoder semantics. A threshold from a different measurement bandwidth must not be shown as an absolute line on the display spectrum.

Acceptance for this evaluation: eight stable receiver paths; selected-slice control and per-slice SQL/mute/frequency isolation; mixed analog WFM presentation; park/resume and persistence; a representative longer receive-only continuity run with actual faults, callback timing and thermal observations reported separately. Historical strict zero-timing/thermal screens are not automatic blockers to this operator evaluation. Development/testing is sequential and proportional; no new routine agent fan-out or broad unrelated suites.

TCI has an existing typed per-slice PCM consumer in source. Eight-client/external-streaming throughput remains to be measured. RTL does not provide Flex DAX/DAX-IQ streams, and no RTL DAX or RadioReference/Broadcastify publishing support is claimed by this amendment. Results and exact candidate identity will follow after focused validation; no merge is requested here.

Nobara SQL/eight evaluation results — 2026-09-28

Final local signed candidate: 48b9afeeb4ba75ec6899c0cfd460e95fa1c89225, tree 751ed3b0db0d392fe46c4de58d0471180419d56d (Ozy311 signature verified). Seven focused native tests passed, including generated-IQ SQL behavior, eight-slot model/restore checks and the 24-case RX applet suite; isolated package startup also passed. The detector-bypass mutation fails. A final fresh-process check found and fixed an additional UI manual-memory restore defect: accepted SQL54 was retained by the backend but the attached Auto UI remembered20. Its regression fails before the fix; final hardware Auto→Off→Manual returns54. Manual scale/defaults are unchanged.

Old/new three-station comparison retained the operator's460.300000FM-N/54,460.275000FM/51,460.175400FM/51 baseline. Old display Auto selected about39 and produced sustained audio; new engine Auto closed the final15-second comparison window on all three. Live RF was sequential and intermittent, not identical recorded IQ or acoustic proof. All eight controls were exercised; H's threshold/margin/tuning/mute changes preserved A–G recipes. Explicit target tuning parked A–G and restored all eight. Final fresh-process restoration retained all eight Auto/manual/filter/monitor recipes.

The ten-minute processing run on ancestor 3d7fea2 delivered28,827,648 frames on every receiver and the speaker mix with zero observed queue drops, mixer faults, withdrawals, nonfinite samples or PCM discontinuities. It recorded394 deadline overruns/175,951 callbacks; mean1.116ms. Process CPU averaged64.63% of one core (~3.23% across20 logical CPUs), RSS859.5→905.6MiB and package-temperature samples70–100°C. This establishes neither a memory plateau nor thermal qualification. The preceding Center-intent parking exercise left H at converter DC (captureDcClear=false), so this is processing continuity, not eight clean RF signals.

The final 48b9afe changes only model manual-memory adoption plus its regression test; detector/DSP/mixer code is unchanged. Its separate60.064-second live interval restored the original460.9MHz capture center with all passbands clear of DC:2,883,584 frames per path, zero observed continuity/fault deltas,392 deadline overruns/17,601 callbacks (mean1.174ms). Both timing results are retained. Both owned runs closed normally; normal profile/launcher hashes were unchanged. The delivery profile keeps only the operator's original three station recipes; capacity is eight and test stations stay in a separate profile.

The local sealed48b9afe package, bound launcher, source bundle and detailed evidence report are retained on Nobara. Bundle SHA-256:41591f452851e7bfc100888aadf8308f579029e3f08ad70750d3de87ebf09054 (base7911324a). This remains an isolated evaluation for maintainer review, with qualified default one. TCI's typed per-slice route was audited, not eight external clients; RTL DAX/DAX-IQ and RadioReference/Broadcastify publishing are not supplied. No PR, merge, public release or Mac/Windows qualification is requested or claimed here.