ULTIMATEEXPERIMENTSCommodore 64 · Work in progress

Optional native media tests

Development note: see current status, build instructions and installation for present qualification. Historical measurements below apply to their stated slice, not every current build.

make test-native-media builds a separate development profile, compiles its firmware model from source and runs27 actual6510/Peddle groups: A/B selection5, cancellation4, online-C10, outputCRC5 and direct Zig3. Disks are deterministic synthetic byte arrays. No supplied DOS images, vendor files, private research folders, network or device access are required.

The optional profile enables DOS64_SAVE_HD_CONTEXT=1 and DOS64_EXPORT_CRC=1 in a fresh staged native project. The normal build keeps both disabled and its existing bytes. make native-media builds/verifies that profile without running models; no Zig, C++ compiler or model resources are needed for that native-only path. Full tests require Python3.9+, make, ca65, ld65, a C++20 clang++ and the project's Zig0.16 toolchain. Set DOS64_ZIG, DOS64_CLANGXX, DOS64_CA65, DOS64_LD65 or DOS64_MAKE for explicit tools. The normal Makefile's Zig lookup works in the standard checkout layout.

Use DOS64_NATIVE_MEDIA_ARGS="--out /an/empty/or/native-media-owned/folder" to select results. Each invocation uses a new run/workspace; native outputs never replace the normal build's main/unit images. build/native-media/RESULT.json is an atomic pointer to the newest fully verified run's eight immutable native products. Sources, compiler resources/tools, command logs, before/after input/product hashes and test reports are retained in that run. Before the final pointer commit, failure or cancellation leaves the previous passed result/products intact and retains run FAILURE/LAST_FAILURE. A host lock cleanup failure is reported honestly; it does not publish success or delete another owner's lock. An interrupt delivered after an atomic commit may leave the complete new result; a partial product set is never published.

Optional builder controls: python3 -B tests/native_media/test_builder.py. They exercise13 host groups including interrupted publication, double failures, failed copy/mkdir/unlink, source and product mutation, owner/path rejection, cache/environment isolation and real native-only/default builds. python3 -B build_support/native_media_tests.py --native-only --default-profile --out build/native-media-default gives a separate default-byte proof. Native-only results have an empty groups object; they are build/memory checks, not executed model tests.

The explicit source inventory stages normal native sources and pinned Peddle sources. Existing tests/config/support headers remain unchanged. Only two dedicated model overrides add ordinary WRITE05, close-failure injection and handle-position observation; their matching canonical base identities are checked. No opaque model library is shipped. New source package files are text; the native build reuses the existing CP437 font resources.

These are SaveAs checkpoints: new owned paths use CREATE_NEW, online C's handle/position/context survives only known-closed alternate-target use, and A/B can read complete360KiB media at REU100000/180000/380000. Source descriptor/policy/mount/content snapshots are captured before borrowing memory and checked again before native success. SaveAs permits readonly or clean media and preserves readonly/dirty/clean-generation/binding state. A fully verified file CRC alone does not establish a selected-media receipt after metadata changes: require current native ACK A5, sav_error0, exact captured media identity/generations and known final closure. CRC00000000 is valid. Partial/canceled CRC is diagnostic.

No CLEAN, CFG publication, backing-path change, persistent Save, Discard/Swap user menu, physical acceptance or V1 completion is claimed. Genuine DOS and Turbo Pascal workflows remain separate acceptance lanes. Model CPU clocks are not physical Ultimate time or an XT performance comparison.

Runtime menu build and composition checkpoint

make runtime-media is the ordinary source-only compiler path for the enabled runtime menu; see its build contract. It produces an immutable runtime/MNU/release-descriptor set with regenerated native/worker aliases. It needs cc65 in addition to the assembler tools, but no DOS files, Zig, C++ compiler or model. It is separate from the older make test-native-media SaveAs regression profile described above.

The complete Insert controller now passes21 actual runtime model groups:9 controller,5 CIA menu,5 composition faults and SaveAs/Finish regressions. Incoming receipt, re-PREPARE, PLAN, Insert WIRE/BIND and native COMMIT execute as linked code; positives do not inject worker-completed facts. Controls include legitimate CRC0, all rotating arenas, first B, late content corruption, failed asset CRC and unknown CLOSE. An independently reproduced stale-result Resume advertisement is fixed by clearing the bridge observation before controller entry. The actual native release guard already blocked unsafe release in the predecessor.

The initial promoted public source build reproduces the reviewed runtime58fa95e2/MNUaa135d2b/REL531c74c3 set. Root independently executes three actual CIA smoke groups from that fresh build using the verified source-built peer model: B SaveAs with Ctarget1, first B Insert with Ctarget2, and controller CRC refusal before file operations. This is synthetic-media model evidence, separate from cold START, real DOS disk revalidation, physical hardware and final V1 acceptance. The21-group source closure and peer reviews are retained in the checkpoint (local development record, not distributed).

The subsequent two-phase clock-admission repair passes13 actual groups and builds publicly as runtime0b523722/MNU4424dfd7. WIRE/Save retry normal transient clock sampling with a256-call limit before touching FILES/context; actual stalled and terminal fault controls retain unchanged C/CFG. This resolves the isolated admission cause; the final CONFIG80 cold consumer rerun and separate broker/bridge pending checks remain open. Clock admission evidence (local development record, not distributed).

FORMAT and combined media runtime

make runtime-media DOS64_RUNTIME_MEDIA_ARGS="--floppy-format" builds the combined FORMAT/media profile. The real runtime is tested directly; full FORMAT1 UNIT currently overflows and is explicitly unsupported. Ordinary UNIT0 and the older selected-profile runtime/UNIT remain byte-identical. A clean public runtime-test target is still pending; archived actual test sources/commands are retained in the composition checkpoint (local development record, not distributed). This limitation is not described as a complete released test workflow.

Root combined runtime806eaab6/MNU1b7086ec passes27 unique actual CPU groups (FORMAT12, admission10, CIA3, new Zig2) and28 host groups. The source-only public build reproduces all9 tested products and checks515 source/tool/resource and474 generated records. Existing generic slot tests now recognize selector8 as FORMAT and reject9; default memory layout/helper checks pass. The actual new Zig test binary is source-compiled and directly executed. Python tests reuse two verified source-built peer models.

A separate genuine CONFIG80 consumer on predecessor0b52 now passes the complete DOS→Save→Insert→B read→C read→Finish after correcting the CIA driver to hold each press/release until an actual scan. The old driver released keys during transient pending polls before scanning; its exact lost path and honest missing-file refusal are retained. This does not count as806e hardware acceptance. All final V1 gates remain open. Evidence (local development record, not distributed).