[repeat98]

FM Synth 0.1.3-experimental moves its SRC page code out of the 16 bytes the stock DELAY reads past its table. It's a card-build fix; the USB package is unchanged.

Draft: waiting for the owner's approval record. Everything below is done and checked except the steps that need it.

What was wrong

The stock DELAY looks up 0x400d1ce0 + 4 * (x >> 6) and keeps the word in the track's delay state. Its table has 508 entries and ends at 0x400d24d0. At a slow tempo with a long TIME the index goes past the end (508 at 51 BPM, measured under the ColdFire port) and reads the next 16 bytes, which are zero on a stock unit.

FM Synth 0.1.2 linked its SRC page at exactly 0x400d24d0 in card builds. In the browser-composed FM-Synth-only image those 16 bytes hold page code (202f0008…), so DELAY at those settings reads code as a coefficient. Not heard on a unit.

What changes

• The page is a DRAM unit, like machine.s and registration.c already were (manifest.py).
• Registration calls the page by its symbol, fm_descriptor, instead of 0x400d24d0. registration.s is the compiler's own output for the changed registration.c (GCC 16.2.0 in the pinned image, same flags as before), byte for byte.
• USB adapter: it used to rewrite those two calls from the fixed address. Now it maps the compiler's jra fm_descriptor back to the absolute jmp it always emitted, so the USB package image stays the one its recipe binds.
• Machine port and recipes: the port's reviewed source hashes and the five machine recipes' source pins are updated. Every recipe still matches its image.
• Version, catalogue, release notes, changelog, README and TESTING.

Tests

• Card image: the 0.1.3 FM-Synth-only MAIN has zeros at 0x400d24d0..0x400d24df; 0.1.2's has code there.
• LCD: all seven captures retaken from the 0.1.3 image with the recorded plan, and from 0.1.2 with the same emulator. All fourteen PNGs are byte-identical to the committed ones, including the SRC page the moved code draws. The 0.1.2 captures' emulator build is gone; these used ot_emu bdbe962c….
• USB package: 0.1.3-experimental.mwrm is byte-identical to 0.1.2's (29,810 bytes). Both stock-free scans pass: longest shared run 11 bytes (limit 12), no 16-byte run. octatrack-module-packages.py --check reports it identical.
• Card packages: only fmpage (now DRAM) and fmregistration (two symbol relocations, 1,408 → 1,464 bytes) change. Every other package changes only its version label.
• Adapters: fm_runtime_tests and machine_composition_tests pass, including the relocation proof (no 0x400d24d0 in the package) and both composition proofs.
• Native parity: the native side of module:verify synth ran (154 selections, cached). Its browser side needs the regenerated documents, so the record isn't written yet.

Not done

• Owner approval record, sdk/synth-build-approval.json, for this version: same waivers as 0.1.2 (current-build hardware, chip worst-case cycles, complete memory bounds). Native source a2d318bb91b4c115dd248c9b439250eb28552e2dc70eb83f967a0e24ad8b7f39. After that: npm run modules:generate, module:verify synth (browser side from cache), the parity record and pr:ready.
• Not rerun: the Octemu MIDI-selector and mono/four-voice audio regressions. The engine and the bundled quantizer are unchanged, and the sample-free fixture those runs used is no longer on the test machine.
• Hardware: none, as for 0.1.2.
• Locally, npm run octatrack:module-packages still stops at Tape Echo's and Euclid's standalone recipes, which were already stale on main after kit 0.7.0. CI skips those builds. The follow-up DELAY-guard PR re-pins them.

Once published, this notifies FM Synth's followers, by email according to their preferences.

🤖 Generated with Claude Code

[repeat98]

@timhastie heads-up on FM Synth: this moves the SRC page code into DRAM on card builds, so the stock DELAY no longer reads it past the end of its table at slow tempos with a long TIME. Your engine and quantizer files are unchanged and the USB package stays byte for byte the same. Tell me if you see a problem with it.

[timhastie]

Thanks for this! I'll fold your SRC page move into FM Synth 0.1.4, credited to you, so #895 can close once 0.1.4 is up.

[repeat98]

Closing: #1146 (FM Synth 0.1.4) moves the SRC page out of the 16 bytes the stock DELAY reads past its table in FM Synth itself, so this fallback isn't needed.

---

Generated by Claude Code