[repeat98]

One beta base for testers this round. Four pieces ship together, with the committed beta base regenerated once from the combined source (6541d7e9):

1. More DSP room: the loader's packet handling runs from the shared window. The DSP receiver's packet operations (WRITE and its read-back, SUM, BIND/UNBIND/BYPASS, BASE, PEEK, the refusal records) now run from the top 512 words of each core's FX2 block 3 (P:0x37e00, P:0x3fe00), copied there from the arena's top at the first frame. The frame path, PROBE, the reply and the engines' ops 10/11 stay in private P. Arena with nothing placed: 4,292 / 4,418 → 4,741 / 4,867 words (+449 per core). Effects still run from private P: code fetched from the window costs twice (measured, below).
2. FX2 swaps keep their own buffer block, and the refusal box speaks plainly. A track swapping one buffered FX2 effect for another was refused MEMORY FULL … NEEDS A BUFFER, FX2 BUFFERS 4/4 whenever its core's four blocks were in use, because the planner reserved the outgoing effect's block. Stock gives every slot a fixed block, and the old effect finishes its sub-block before the new one's init, so the slot now keeps its block. A remaining refusal reads NO ROOM / the effect / FX2 ON T1-4 FULL / CHANGE ONE FIRST.
3. The fix/base-oct28 bundle (merged here at 8dd674f7): POLY8 0.3.2 (card chooser scroll, [vector] Vector & Poly8 together, Vector missing from machine selector window #842 [synth] A minor bug #941 [configuration] FM Synth not showing on SRC list #942), module machines own their SRC SETUP page + VECTOR 0.2.5 ([recorder-loop-fix] Vector show just 1 SRC menu #972), SRC SETUP encoder E (SETUP_ALL) and the LIST flag (this contains Fix: Octatrack base, SRC SETUP's encoder E and the double-tap window for module machines (SETUP_ALL, LIST flags) #943's head), FLEX tracks no machine owns render unguarded (Octatrack base: frame guard skips stock FLEX tracks while FM Synth is installed #999).
4. The Python half of Modwerk's own firmware codec (own-firmware-codec-py @ abc7055b), which the owner chose to ride on today's base: sdk/firmware_codec and use_elekloader() for every Octatrack tool. Its changelog is merged with Improvement: Modwerk's own firmware section codec and ColdFire decoder #1016's.

What players see

• About 450 more words of effect code per DSP core: finding 53's heavy stock Part (LO-FI, FILTER, DJ EQUALIZER, CHORUS + DARK, SPRING, PLATE REV, FLANGER on T5-T8) now loads whole, and the project's module set on the test unit (BBD, Tape Echo, Mini Verb, E-Verb, Air Chorus) installs where origin/main refused it ("Mini Verb doesn't fit on either DSP core").
• Owner-accepted limit: Air Chorus and BBD each declare a whole 16K FX2 buffer, so at most two of tracks 1-4 (and two of 5-8) run them on FX2 at once (were three); E-Verb, Mini Verb and every stock effect keep all their blocks.
• FX2 swaps no longer fail on a full core; the one remaining refusal is in words a player understands.

How I checked it

• Measured on the owner's MKII (core 0, dev builds, --dsp-burn-window): a 2,000-cycle loop cost 100.4 idle passes from private P and 203.5 from P:0x36000.
• Host tests: test_recipe.py (39, incl. new test_window_buffers.c: block-3 rule, wide-first planning, own-block swaps) and the resident placement/board suites; the resident base patch was 3-way merged onto the split receiver and its pins re-set (it reproduces the merged sources exactly).
• Base: npm run octatrack:base -- --beta --write → 6541d7e9, every flash-safety and browser = native check passed; a control run on clean origin/main reproduced the committed base byte for byte first.
• Packages: npm run octatrack:module-packages -- --check: all 27 published USB packages rebuild byte-identical with the merged tooling (POLY8 stays at 0.3.1 and Play Modes at 0.1.1, both declared lags). VECTOR, FM Synth and Scale Quantizer's machine recipes were recaptured; only their input pins changed (image hashes the same). Isolated card-module build (build-modules-isolated.sh, pinned image) imported at 03006a5b; Tape Echo's and Euclid's recipes pin the new assets. Analog BD's 212 native composition proofs re-exported identical, only moduleSourceTreeSha256 restamped.
• On the owner's MKII (RAM boots, no flash): every build's boot uploads ran through the window half on both cores (0 rejected, 0 errors, 0 stalls). Base 6541d7e9: self-test passed; Sidechain, Analog BD, E-Verb and BBD synced with a passing self-test. Its development twin played an 80 s session (PLAY, picks of DARK REV, LO-FI, Air Chorus, SPRING and PLATE REV and E-Verb while playing, Part 2 and back, bank B and back) recorded over 20 USB channels, the same as a twin of origin/main: MAIN 0 dropouts and 0 clicks on both, the largest jump around each action within 2 %, 0 rejected packets, 0 rollbacks. A 2.2 ms gap on T4 about 1.3 s after picking DARK REV there appears on both builds (not this change). Earlier twin: finding 53's Part loaded (4,665 of 4,741 words), a third Air Chorus on T1-4 refused with the new box, E-Verb took block 3.
• npm run pr:ready: Ready.
• Not tested: a flashed base and power-on restore of a kept set; long churn runs (owner: test like a player instead).

Owner decisions recorded here

• The two-per-core Air Chorus/BBD limit for the arena room (28 Oct).
• POLY8 0.3.2's owner-approved update re-approved for the later source of f1a25004 (clamp bytes in both manifest forms; USB package unchanged): sdk/poly8-build-approval.json.
• sdk/tests/test_firmware_codec.py names scripts/elekloader-vendor.ts as text-only: since Improvement: Modwerk's own firmware section codec and ColdFire decoder #1016 it lists the replaced Python files only to drop their NOTICE credits, and never runs the checkout. This is the codec branch's own new test meeting Improvement: Modwerk's own firmware section codec and ColdFire decoder #1016, not a change to a gate on main.

What the guides lacked or got wrong, and where I fixed it

• Nothing said that code fetched from the shared DSP window costs twice private P, or that Modwerk's base keeps stock's boot loops (which zero the whole window): docs/OCTATRACK_HARDWARE_TRAPS.md (the window bullet, rewritten, no growth), with the dated record in a new docs/OCTATRACK_DSP_WINDOW.md (listed in docs/README.md).
• docs/module-guides/machines.md (from the bundle) linked the renamed EVENTS.md: now HOOKS.md.
• npm run device -- meter 0|1 asked for a report length current bases refuse, so it never read: fixed in scripts/device.mjs.

Related to #999, #972, #842, #941, #942. Contains #943 (close it after this merges if it stays open).

🤖 Generated with Claude Code