refactor(radio): flatten LS state and only save frozen state per FM

[raphaelcoeffic]

Summary

Flatten per-FM logical switch state (lswFm[MAX_FLIGHT_MODES]) to a single global context (lswCtx[MAX_LOGICAL_SWITCHES]), saving ~2 KB RAM and making logicalSwitchesTimerTick 9× faster. Clarify the relationship between evalLogicalSwitches and the mixer fade loop. Encapsulate the Lua sticky switch buffer behind a proper API.

Background

Per-FM logical switch state was introduced in opentx/opentx#1119 (May 2014) to fix a reentrancy bug (opentx/opentx#1039) and give each flight mode independent LS evaluation during fade transitions. One month later, commit f3bc8cb9a4 (fixing opentx/opentx#1345) gated evalLogicalSwitches() behind if (tick10ms), which is zero for non-current FMs during fades. This made per-FM LS evaluation dead code — only the active FM's context is ever evaluated, while non-active FM contexts carry stale state from the moment of transition.

The per-FM state has been inert for 12 years.

Changes

Commit 1: test(radio): prove per-FM logical switch state is redundant

• Test suite proving that timer/sticky/edge lastValue is always identical across FMs
• Test proving that non-current FM's LS state is stale during fade transitions (never independently evaluated)

Commit 2: refactor(radio): flatten per-FM logical switch state to single global context

• Flatten lswFm[MAX_FLIGHT_MODES] → lswCtx[MAX_LOGICAL_SWITCHES] (single global array)
• Hide LogicalSwitchContext struct as implementation detail in switches.cpp
• Collapse logicalSwitchesTimerTick from FM×LS nested loop to single LS loop
• Replace logicalSwitchesCopyState (full context copy on FM transition) with lswFreezeState (bitmap snapshot of LS state bits into uint32_t array)
• Add mixerActiveFlightMode to distinguish active FM from temporarily assigned mixerCurrentFlightMode during fade evaluation
• getSwitch() reads from frozen bitmap for fading-out FMs, live context for active FM — preserving existing fade cross-fade behavior

Commit 3: refactor(radio): move evalLogicalSwitches out of evalFlightModeMixes

• Move evalLogicalSwitches() from inside evalFlightModeMixes to evalMixes, called once before the fade loop
• Remove the isCurrentFlightmode parameter (was always true since the tick10ms guard prevented the false path for 12 years)
• Expand the fade loop ternaries into an explicit if/else for clarity
• This makes the intent explicit: LS evaluation happens once per mixer cycle for the active FM, not per-FM during fades

Commit 4: refactor(radio): encapsulate sticky switch buffer behind lswSetStickySwitch

• Add lswSetStickySwitch(idx, value) API in switches.cpp; Lua calls this instead of encoding messages directly
• Replace CircularBuffer<uint8_t> (bit-packed, limited to 64 LS) with a lock-free SPSC ring buffer using std::atomic for correct cross-task ordering (GUI task → mixer task)
• Remove CircularBuffer class from edgetx_helpers.h (no remaining users)
• Remove MAX_LOGICAL_SWITCHES == 64 compile-time assertions

Preserving fade transition behavior

During FM fade transitions, the mixer evaluates mixes for multiple FMs per cycle. Fading-out FMs previously saw their (stale) LS state from lswFm[]. To preserve this behavior, a uint32_t bitmap per FM captures the frozen LS state at transition time. This replaces ~2 KB of per-FM context with 72 bytes of bitmaps + 1 byte for mixerActiveFlightMode.

Test plan

• Existing test suite proves per-FM state redundancy
• New test (perFmLsStateNotEvaluatedDuringFade) verifies fade behavior
• Full gtests-radio suite passes on TX16S, NB4P, GX12
• Flight test with FM fade transitions and LS-conditioned mixes

🤖 Generated with Claude Code

[raphaelcoeffic]

This works as before with @philmoz 's example model from #7199:
tx15.ls.fm.test.etx.zip

[raphaelcoeffic]

Some test results with E-Soar Plus template:

• RM Boxer without FM transition:

main:                Tmix max    1.20ms
flatten-ls-state:    Tmix max    0.93ms (77.5% of baseline)

• RM Boxer with FM transitions:

main:                Tmix max    1.90ms
flatten-ls-state:    Tmix max    1.57ms (82.6% of baseline)

• RM TX15 without FM transition:

main:                Tmix max    0.52ms
flatten-ls-state:    Tmix max    0.55ms (105.7% of baseline)

• RM TX15 with FM transitions:

main:                Tmix max    0.74ms
flatten-ls-state:    Tmix max    0.70ms (94.6% of baseline)

[RC-SOAR]

Just resumed testing with RM GX12

• Flashed pre-3.0.0-PR7219 eb2bd7b7
• Started Companion 3.0.0--main Mar 18 2026
• Opened .etx file
• Transferred setups to tx.

The setups transferred successfully to the tx, however there were a couple of glitches:
(1) initially, the switches didn't work (failed to respond in Sys>Hardware>Debug Keys menu). Switches recovered after reboot.
(2) Most if not all SF's were disabled (repeat of old issue?)

I fixed the above, then tested a complex (F3F) setup using the mixer monitor - seems to work but really need to visualise as it's impractical to assemble the models each time. I'm going to try a quick and dirty port of my visualiser script to the TX16S.

[raphaelcoeffic]

Just resumed testing with RM GX12

• Flashed pre-3.0.0-PR7219 eb2bd7b7
• Started Companion 3.0.0--main Mar 18 2026
• Opened .etx file

Could you please upload the models prior to conversion using Companion? I'd like to take potential issues with current Companion out of the equation.

Also, it seems the models are using LUA scripts which are not included (ex: /SCRIPTS/MIXES/snp500.lua).

• Transferred setups to tx.

Ideally, and to really exclude potential pre-existing issues with current main branch, you would:

• first flash latest nightly release
• transfer models
• verify everything is ok
• then flash the firmware from this PR
• verify things still work

Alternatively, you could use 2.12.0 instead of nightly.

The setups transferred successfully to the tx, however there were a couple of glitches:
(1) initially, the switches didn't work (failed to respond in Sys>Hardware>Debug Keys menu). Switches recovered after reboot.

That would be the radio settings. I'll have a look specifically there. So far, I could not reproduce it here.
Ideally, I'd need a copy of /RADIO/radio.yml before and after loading it into the radio. That might show something regarding the corruption. So far, @philmoz reported something like that as well (was calibration data in his case), but I couldn't reproduce it.

(2) Most if not all SF's were disabled (repeat of old issue?)

It seems to happen as well if the same models are loaded with current nightly release. So possibly not something introduced by this PR or #7199.

Loading your .etx file in latest Companion built from main branch shows most of the SFs as disabled (all but SF2 & SF3 on "FS5_620"). This might have been introduced by the conversion done with Companion. Companion 2.12.0 might be a better choice to do the initial conversion.

I fixed the above, then tested a complex (F3F) setup using the mixer monitor - seems to work but really need to visualise as it's impractical to assemble the models each time. I'm going to try a quick and dirty port of my visualiser script to the TX16S.

Is there anything that would help you testing? Would it be easier to test using the Companion simulator?

[3djc]

@RC-SOAR Mike, the GX12 issue is different from this PR, as SA and SD have moved to switches to custom switches, you must be experiencing side effects about that. Could we have you initial model ?

[RC-SOAR]

Loading your .etx file in latest Companion built from main branch shows most of the SFs as disabled (all but SF2 & SF3 on "FS5_620"). This might have been introduced by the conversion done with Companion. Companion 2.12.0 might be a better choice to do the initial conversion.

Aha - it's an old issue! I've just checked - the SF's were already disabled in the 2.10 .etx file before import into Companion 3. The reason being that the 2.10 .etx had itself been converted from an OpenTX file, and SF's were disabled during that conversion - I seem to remember this was a legacy issue which was fixed.

Is there anything that would help you testing? Would it be easier to test using the Companion simulator?

I can test using the actual hardware, with my visualiser script (now supports colour radios) - it would be easier than using Companion.

Ideally, and to really exclude potential pre-existing issues with current main branch, you would:

• first flash latest nightly release 2.12.0
• transfer models
• verify everything is ok
• then flash the firmware from this PR
• verify things still work

All noted. I will make a start tomorrow. Is it sufficient to test using a single target? (I can test on a TX16S Mk1, TX16S Mk3 and/or TX15.)

The setups transferred successfully to the tx, however there were a couple of glitches:
(1) initially, the switches didn't work (failed to respond in Sys>Hardware>Debug Keys menu). Switches recovered after reboot.

That would be the radio settings. I'll have a look specifically there. So far, I could not reproduce it here. Ideally, I'd need a copy of /RADIO/radio.yml before and after loading it into the radio. That might show something regarding the corruption. So far, @philmoz reported something like that as well (was calibration data in his case), but I couldn't reproduce it.

Just to clarify, the switch failure occurred after exit from the bootloader "flash firmware" screen. The problem was at a low level, switches were not responding in the Hardware Debug menu. After reboot all was normal. The switch assignments were correct for the GX12 before import (SA/SD not used). I'll see if I can reproduce the issue using a more stable test setup.

[raphaelcoeffic]

All noted. I will make a start tomorrow. Is it sufficient to test using a single target? (I can test on a TX16S Mk1, TX16S Mk3 and/or TX15.)

Yes. The goal is really to check if FM transitions are the same as before. The change in this PR is really only about the mixer and per-FM LS state. It works the same on all targets.

Just to clarify, the switch failure occurred after exit from the bootloader "flash firmware" screen. The problem was at a low level, switches were not responding in the Hardware Debug menu. After reboot all was normal. The switch assignments were correct for the GX12 before import (SA/SD not used). I'll see if I can reproduce the issue using a more stable test setup.

This might have been caused by the previous version which had a use-after-free memory issue, which can cause uncontrolled memory accesses and thus break basically anything. That should be fixed now.