[npp1993]

Three Octatrack base fixes for module machines (Analog BD, FM Synth and the others that run on a FLEX track), combined on one branch as agreed, plus the USB bridge fix they were tested with. Author: Nathaniel Pavel (npp1993).

• What was wrong or hard:
  1. After a power cycle, PLAY on MODULES LOADED left Analog BD silent until the next stop: an on-demand engine was placed only while the sequencer was stopped and settled, a controller RESET went one 24-word packet a 60 Hz tick (about 2.6 s a core for Analog BD), and the restore said MODULES LOADED before any engine ran.
  2. FLEX ↔ module machine switch: the track's live parameters stayed on the old machine's values until SOURCE CLEAR, so a machine chosen from FLEX could stay silent and FLEX chosen back played the machine's settings.
  3. After a boot, a restored module-machine track showed "SRC > FLEX" until it was tapped.
  4. USB bridge: with another Elektron device on USB (an Analog Rytm on Overbridge) the bridge opened the wrong device and every request to the Octatrack failed.
• What this changes:
  1. An engine is placed while playing when it fits in free words (dl_engine_claim commit 1); trims, evacuations and unplacing still wait for a stopped, settled unit, and nothing is placed while a pattern change is queued. RESET jobs carry up to eight packets. MODULES LOADED waits for modwerk_controller_settling() (0.5 s at least, 10 s at most; a closed core, a lost core state, a key bus that cannot come up or a stuck engine ends the wait). A packaged feature counts as used once a track of the bank's Parts is signed for its module's machine (modwerk_machine_bank_modules), not only from a track's first trig. bind_live keeps a module's retry when a later module in the same pass binds. Owner decision asked: this relaxes the "no placement while playing" rule in docs/OCTATRACK_RESIDENT_ENGINES_ON_DEMAND.md (the doc is updated to say what still waits for a stop).
  2. A FLEX ↔ module-machine switch reloads the track's live parameters from its Part, as stock does when a switch changes the machine type (machine_registry.c). STATIC and the other stock machines are unchanged.
  3. runtime.c redraws the track header after the modules restore (the stock header drawer at 0x4004d780).
  4. scripts/octatrack-usb-bridge.py opens the device that offers the Modwerk interface.
     Changelog: changelog/2026-10-27-engines-play-after-boot.md, 2026-10-27-flex-machine-switch-live.md, 2026-10-28-boot-src-label.md, 2026-10-27-usb-bridge-elektron-device.md. The beta base is regenerated (base-binding.json, the engine assets, the base-bound packages).
• How I checked it: npm run pr:ready (on this Mac only test_published_module_packages_rebuild_byte_for_byte fails, on clean main too: the Tape Echo standalone recipe; CI has the inputs). Host: test_resident_engines_board.py modes 1, 13, 46, 54, 64; test_resident_transport.c, test_modset.c, test_machine_registry.c, test_runtime.c, sdk/tests/test_usb_bridge.py. On the MKII (OS 1.40C), recorded over USB, see docs/verification/2026-10-27-engines-play-after-boot.md: on a development twin, a FLEX → Analog BD switch while playing was placed on both cores within a second and bound in 1.8 s; a bank change into an Analog BD bank while playing placed it in about 1 s with T2 and T3 bit-exact across the change; on a flashed personal base (14 modules), MODULES LOADED then PLAY sounded Analog BD on the first step, and T1 showed SRC → ANALOG BD without a tap on every boot; FLEX/STATIC → Analog BD played without SOURCE CLEAR in a blank project. On the final build (ccc53692 = this branch plus Analog BD 0.1.6, Nathaniel's own project, 28 October) all three pass again: SRC label, first trig after MODULES LOADED, FLEX ↔ Analog BD switch. Not fixed here, and older than this change: a DSP halt in the first minute after a cold boot (DSP STOPPED, residency off for the boot). It happened with FUNC held (no modules restored), so it is in the base; the five stall records are decoded in the verification record and point at a lost frame-interrupt edge, not at these changes. Open too: Ringmod SC refused at the boot restore on some bases (RUNTIME_BUSY), a first-trig stutter after Analog BD → FLEX, and an FM Synth switch untested.
• What the guides lacked or got wrong, and where I fixed it: docs/OCTATRACK_RESIDENT_ENGINES_ON_DEMAND.md (placement while playing; an edited elekloader file the resident patch touches needs its pin and patch context updated), docs/OCTATRACK_DRIVING.md (the bridge and a second Elektron device), docs/OCTATRACK_BETA_TEST_PLAN.md, docs/module-guides/machines.md (the switch reload), the two verification records.

Please merge #944 (Analog BD 0.1.6) first if you can; this one then regenerates its base on top. Either order works, whichever lands second regenerates the base.

🤖 Generated with Claude Code

[repeat98]

Thanks, this is a big step for module machines after a boot, and the verification record makes it easy to follow. A few things before it can go in:

1. The last merge of main dropped a line from Fix: Octatrack card builds, too many module voices no longer stall the unit (frame guard) #1030. In docs/module-guides/machines.md, "A bounded render" now says "card builds have no guard" again. Main says card builds' core logger does the same (sdk/runtime/logging/frame_guard.c) and logs the dropouts as GRD. Please take main's sentence back; your new "Switching sounds right at once" item above it is fine.
2. The bank scan runs on every tick, with interrupts masked. use_tick() calls modwerk_machine_bank_modules() from inside modwerk_controller_tick()'s modwerk_machine_lock() section. It scans 8 Parts × 8 tracks for each machine, for as long as a held packaged feature isn't seen yet. When Analog BD is installed but this bank doesn't use it, "not seen yet" stays true, so the scan runs every tick for the whole session. Please scan only when the bank or a Part changes (or a machine is assigned), or every N ticks, or outside the lock. Or measure it with the guard (npm run device -- guard) in a busy pattern with Analog BD installed but unused, and put the number in the record.
3. The SRC label changelog (changelog/2026-10-28-boot-src-label.md) says "not yet on a unit". Your record has it passing on every boot on the flashed base and on ccc53692, so please say that.
4. POLY8 and VECTOR. They have their own assign, so mw_assign returns before relive() when you switch FLEX → POLY8/VECTOR. Switching back to FLEX now reloads the live parameters. That's probably right, but it changes how those two leave a track, and nobody has listened to it yet. Either check POLY8 → FLEX and VECTOR → FLEX on the unit, or say in the machine-switch changelog which machines it covers ("Analog BD, FM Synth and the others" reads wider than it is, and FM Synth hasn't been tried yet).
5. Small: the .gitignore change (node_modules/ → node_modules) isn't part of this fix. Please drop it, or tell me why it's needed.

On placing engines while playing: that's fine with me, as long as point 2 keeps the bank scan off every tick. Once #944 or #1033 lands first, this needs main merged in and the base regenerated (and #1033 also adds a test mode 64, so whichever goes second renumbers).

[repeat98]

Thanks, that covers all five points. The bank scan now runs only when something changes, or every 64th tick, and the docs and changelogs read right.

One new thing: module-paths-windows fails in the board test with undefined reference to modwerk_machine_assigned. MinGW doesn't resolve the weak, uninitialised volatile uint32_t modwerk_machine_assigned in controller_board.c, so the fixture's extern has nothing to link to. It's the only weak variable in the tree; the weak functions are fine. On PE/COFF a weak tentative definition stays a common symbol, so it never gets defined. Giving it an initialiser (__attribute__((weak)) volatile uint32_t modwerk_machine_assigned = 0;) should be enough. Or define it in the board fixture, as the fixture already does for modwerk_controller_lease_seen, if controller_board.c is always linked beside machine_registry.c. Linux passes, so this only affects the Windows link.

Your flashed runs came before the scan change. On the base you test #1032 with, could you also do one cold boot with Analog BD on T1 and press PLAY as soon as MODULES LOADED shows? For what it's worth, I merged #1032 into this branch locally: the sources merge cleanly and the machine-registry and board host tests pass together. Only the generated base files need regenerating.

[repeat98]

Thanks, Windows is green now. One question about how it's fixed: controller_board.c now defines volatile uint32_t modwerk_machine_assigned; itself, and machine_registry.c:303 defines it too. Both go into the base image, compiled with -fno-common. So either the base's link merges them, or it keeps two copies. If it keeps two, use_tick() reads a counter that mw_assign never increments, and on a unit the rescan on assignment never fires; only the every-64-ticks path does. The board test can't show which, because the fixture increments the counter itself.

Could you check in the built base that there's one modwerk_machine_assigned, the registry's (for example from the link map, or a symbol listing of the ColdFire objects)? If there are two, make the one in controller_board.c an extern, since it always links beside the registry in the base, and give the board fixture its own definition.