simplex-chat/scripts/nix/README.md

Haskell builds with Nix

The compiler and all cross-compilers come from the haskell.nix input pinned in flake.lock. The ghc914-sh compiler is the stable-haskell fork, not the stock GHC 9.14 release from GHCup. Use the same source checkout and lock file for every platform; Android ARMv7 does not require a separate branch or tag.

compiler-selection.nix and compiler-cache-overlay.nix match Hydra's compiler build context, including auxiliary iserv projects, so the pinned compiler can use its existing cache. Application dependencies keep the project's current packages.

On Linux x86_64, native-rts-project.nix patches the final compiler's RTS ELF loader to honor section alignment even when its file offset is unaligned. The stage2 RTS package patch preserves the configured source, stage1 compiler, plan, and package IDs, but rebuilds the final compiler and its dependent slices. Native, auxiliary, and cross compiler selection share this patched compiler. The ELF fix does not patch target RTS sources or target iserv runtimes, including musl.

The focused regression runs before the native x86_64 Linux CI build. To run it from this checkout, use nix build .#checks.x86_64-linux.ghc-elf-alignment -L. It loads the same 8192-byte C object with aligned and deliberately packed section file offsets, requiring both to retain 64-byte memory alignment and their data. It is a flake check, separate from the 21 existing Hydra artifact outputs.

On Intel Darwin, darwin-info-table-project.nix patches the GHC x86 backend in both compiler stages. It marks info-table entry symbols as Mach-O alternate entries, keeping each table and its code in the same linker subsection. The observed Setup GC abort reads closure type zero because LLD inserts eight bytes between a valid FUN_STATIC table and its entry. Relinking the actual cached directory object reproduces that layout error; changing only alternate-entry symbol flags preserves the table's original 24-byte distance and closure type 14. The patched compiler rebuild and full Intel outputs still require CI validation.

Run nix build .#checks.x86_64-darwin.ghc-macho-info-tables -L to compile a Haskell fixture with the selected compiler and check its table placement after a real LLD link. CI runs this check after the Intel artifact builds. It checks object metadata, table distances, and closure types without executing the fixture.

Intel Darwin projects also declare bitvec's Clang runtime dependency before planning. GHC's object linker needs the selected C compiler's libclang_rt.osx.a for __builtin_cpu_supports; its CPU model is otherwise unresolved during Template Haskell. The SIMD implementation remains enabled. The patch declares both ordinary and static C libraries for TH and rolled archives. The patched local dependency keeps its tests and benchmarks disabled. Its package-local Nix dependency map treats clang_rt.osx as provided by that selected compiler resource directory; all other library mappings stay intact. The resolver regression exercises the pinned HN plan/package/component modules:

nix eval --impure --json --file scripts/nix/check-darwin-bitvec-resolver.nix \
  --apply 'test: let f = builtins.getFlake (toString ./.); in test { lib = f.inputs.haskellNix.inputs.nixpkgs.lib; haskellNix = f.inputs.haskellNix; }'

CI checks this platform scope before the Intel builds with nix eval --json --file scripts/nix/check-darwin-bitvec-runtime.nix.

Cross projects add ghc-internal to the Cabal dependencies of the pinned composition-1.0.2.2 source. This preserves its deliberate avoidance of base while making GHC 9.14's wired-in types available in isolated target package databases. The patch runs before planning and source packaging. Native builds and the Mac-to-iOS conversion projects keep the original package metadata.

Musl projects declare OpenSSL in Cabal's extra-libraries-static field for direct-sqlcipher and simplexmq. GHC's fully static link path reads that field, so an ordinary extra-libraries entry alone does not supply libcrypto. These source metadata patches run before planning and require Cabal specification 3.8.

Linux-built musl ghc914-sh iserv projects use a package-specific executable-static: False policy before Cabal v2 planning. The build-time interpreter needs musl's dynamic C loader for Template Haskell dependencies; a fully static musl executable has a dlopen stub. Its Haskell library ways and profiling selection stay unchanged, and the final CLI remains fully static. An i386-only preplanning source patch omits the private __aio_close symbol registration: shared musl exposes close, which calls its hidden AIO hook internally, rather than exporting that implementation detail. The scope check uses the real Nix module fixed point in the Linux static CI group, including pkgs supplied through _module.args. Repeat it with nix eval --impure --json --file scripts/nix/check-musl-iserv-scope.nix --apply 'f: f { lib = (builtins.getFlake (toString ./.)).inputs.nixpkgs.lib; }'.

The i386 musl target RTS also handles duplicate hidden global function signatures at the start of a sole executable COMDAT member. A known function size must cover the whole member. GCC's __x86.get_pc_thunk.* PIC helpers use zero symbol size (unknown extent); that case permits only the signature plus local untyped labels and section metadata in the member. Additional typed or nonlocal definitions remain errors. Coalescing is checked against the current symbol owner during discovery and loading; it does not persist weak-symbol status or suppress ordinary strong conflicts. Multi-member groups remain outside this patch's scope. The source patch applies to application and interpreter projects, including profiling. Native compilers and other cross targets keep their existing RTS sources.

Run the native x86_64 Linux regression with nix build .#checks.x86_64-linux.ghc-elf-comdat-functions --no-link -L --max-jobs 1 --timeout 7200. It links the same configured RTS sources with and without the patch, verifies the replacement objects in both linker maps, checks relocations in both load orders, and rejects strong conflicts and stale archive-owner decisions. Each probe has a 15-second bound. Module scope is checked separately by scripts/nix/check-musl-rts-scope.nix using the same invocation as the iserv scope check above.

Windows patches Basement's INVALID_HANDLE_VALUE expression to INT_PTR, so hsc2hs receives the signed pointer-sized integer expected by the Haskell API. The generated constant converts through Haskell Int with Prelude.fromIntegral; Basement's RebindableSyntax has no HasNegation IntPtr instance. Windows also plans the already selected crypton-x509-system-1.6.8 as a local patched source, so its Cabal metadata names crypt32 before source packaging. The pinned v2 builder otherwise replaces a package's patched Cabal file with its original description override and copies that override into the slice index. The source producer applies the existing haskell.nix patch once; its later package patch list is empty to prevent a second application.

Linux-built Win64 ghc914-sh v2 projects patch only the RTS DLL search-path caller, including normal and profiling interpreters. It resolves relative paths once, retains successful namespace conversion for long paths, and uses the resolved absolute path if conversion returns NULL. Invalid paths fail without passing NULL to AddDllDirectory; its opaque cookie and Win32 errors survive cleanup. Other filesystem helpers and native compilers keep their sources. The native caller regression uses scripted APIs, not Wine or a target ABI test: python3 scripts/nix/test-windows-dll-search-path.py --rts-source /path/to/configured-ghc/rts. scripts/nix/check-windows-rts-scope.nix checks the real module fixed point using the same pinned-lib invocation as the musl scope check above. Production Win64 RTS compilation and the canonical Wine TH gate remain runtime qualification.

Run these commands from the repository root:

make help
make nix-list                  # list this host's existing Hydra outputs
make nix-native                # native library and CLI
make nix-android               # ARM64 + ARMv7 libraries, on x86_64-linux
make nix-windows               # Windows library and CLI, on x86_64-linux
make nix-linux-static          # x86_64, i686, and ARM64 static Linux CLIs
make nix-build                 # all existing outputs for this build host

Each command builds this checkout, without cloning another branch. Output links are placed under result-nix/<build-system>/<package>. ZIP filenames inside the outputs retain the existing Hydra names. For example, the ARMv7 client library is result-nix/x86_64-linux/armv7a-android:lib:simplex-chat/pkg-armv7a-android-libsimplex.zip. Nix includes tracked working-tree changes; add new source files to Git before building so they are included.

For a remote builder, select its build system explicitly. Nix must already have that builder configured:

make nix-ios NIX_SYSTEM=aarch64-darwin
make nix-darwin-libs NIX_SYSTEM=x86_64-darwin
make nix-build NIX_SYSTEM=aarch64-linux

nix-ios produces the Haskell static-library ZIP. It does not run Xcode, sign an iOS application, or upload to TestFlight. ARM64 device libraries and Intel simulator libraries retain their separate outputs. The existing iOS preparation scripts retag ARM64 archives with mac2ios -s when an Apple Silicon simulator is needed.

The Cabal v2 builder stores ordinary package archives in its staged store. link-mobile-library.sh relinks the selected package and its dependencies before that store is pruned. iOS ZIPs contain both the plain libHSsimplex-chat-*.a and the rolled libHSsimplex-chat-*-ghc*.a used by update-pbxproj.sh. Android and Windows clients link the threaded RTS into libsimplex.so or libsimplex.dll; Android's separate libsupport.so resolves RTS symbols from the client library. Android client export flags use the ordered chat-exports.txt manifest. The desktop library scripts check that manifest and libsimplex.dll.def against the Haskell exports before building; update the manifest when adding an export. Run scripts/desktop/test-exports.sh to check all three guards without compiling. ARM64 Android libraries retain the 16 KiB linker alignment. These steps run in the derivation's build/install phases because v2 component overrides do not run the older postBuild and postInstall hooks.

Mobile builds keep the RTS's small-address-space setting. On Darwin, the final link uses ABI-checked replacement RTS registrations, plus a replacement text on ARM64. These boot libraries and all ARM64 mobile app C/C++ code target the Apple A7 instruction baseline before mac2ios retags the archives. These mobile replacements use the cached ARM64 Darwin compiler; the Intel compiler receives the info-table patch described above. Android applies the RTS source patch when building its boot libraries. The Android overlay also supplies GNU iconv headers and a static PIC archive to ghc-internal, including the separate iserv projects. The ghc-internal system dependency propagates GNU iconv to interpreter consumers as well. Both Android interpreters register the GNU iconv function addresses for runtime object loading. The ARM64 table uses the selected iconv.h to distinguish GNU and Bionic, preserving plain lookup names alongside GNU names when GNU supplies the definitions. Actual API26/API35 C controls check these names against the target archive and the ghc-internal consumer; the Bionic API35 object stays byte-identical. ARM32 source is patched before Cabal planning, removing false missing-function stubs that shadow real static Bionic startup routines. Android links libffi through the RTS package dependency. Android preplanning patches remove the separate pthread library declarations from crypton and direct-sqlcipher. Bionic supplies these functions in libc; the NDK has no libpthread. Other targets retain their original declarations, and compiler -pthread flags and the threaded RTS remain enabled. The local crypton override retains its dependency-only scope; it does not add the package's tests or benchmarks as planning goals. Repeat the scope regression with nix eval --json --file scripts/nix/check-android-pthread-scope.nix. The minimum Android API stays at 26. For Android below API28, SplitMix's initializer uses Bionic's existing arc4random_buf rather than the newer getentropy API. The dependency patch preserves the OS entropy source; API28 and later keep the original initializer. The ARM32 interpreter registers the real arc4random_buf address for loaded objects on that route; the ARM64 interpreter already contains that registration. The ARM64 interpreter registers getentropy and getrandom only from API28, matching the NDK headers and versioned shared-library exports. API26 builds omit those entries; API35 retains them. The interpreter source patch applies to both Android ABIs; only ARM32 replaces the emulator. The Android CI group runs the pure overlay-scope regression with nix eval --json --file scripts/nix/check-android-iserv-scope.nix.

Android ARM32's QEMU wrapper preserves the guest's PER_LINUX32 execution domain when an ARM64 host kernel rejects that domain. Upper personality flags still go through the host, and queries and setters retain the previous guest value. This QEMU patch applies only to the Android ARM32 emulator; the other emulator targets stay unchanged. Android's libc initializes this state again at startup after an exec.

A standalone C regression checks startup, full personality values, the host's UNAME26 effect, and independent thread/fork state. Run it from this checkout with configured Linux builders; Nix supplies the pinned Android API26 NDK and patched QEMU. The NDK compiles the tiny payload on x86_64 Linux; QEMU then runs on the selected Linux architecture. For example, on an Apple Silicon Mac whose Linux builder supports both x86_64 and ARM64:

nix build --impure --file scripts/nix/check-android-personality.nix \
  --argstr system aarch64-linux --max-jobs 1 --cores 1 --print-build-logs

Use --argstr system x86_64-linux for an x86_64 Linux builder. This optional check does not add a Hydra output. It checks C startup and personality handling; it does not qualify GHC/base under Rosetta, cross-TH, or Android device behavior. The static C fixture supplies the ARM TLS alignment input added to newer Android CRTs; its minimum API stays at 26. On 2026-10-15 this regression passed on native ARM64 Linux, native x86_64 Linux, and x86_64 Linux through Rosetta on an Apple Silicon Mac.

NIX_ARGS passes additional options to nix build, and NIX_OUT_DIR changes the output-link directory. Each derivation has a two-hour timeout by default; set NIX_BUILD_TIMEOUT in seconds to change it. For a wall-clock limit on a local batch, use timeout from GNU coreutils to include environment setup in the limit:

timeout 6h nix develop -c make nix-build
make nix-native NIX_ARGS=--dry-run

.github/workflows/nix.yml runs these same outputs on Linux ARM64/x86_64 and Darwin ARM64/x86_64, with a six-hour limit per build job. Its seven build jobs cover all 21 outputs once: x86_64 Linux runs native, android, windows, and linux-static separately; ARM64 Linux runs native; both Darwin hosts run all. A failed native Linux build therefore leaves the cross-build jobs running. PRs, pushes, and default manual runs retain that complete matrix. A manual dispatch can select the existing three linux-static outputs with gh workflow run nix.yml --ref BRANCH -f group=linux-static, or the four Android outputs with gh workflow run nix.yml --ref BRANCH -f group=android, or the two Windows outputs with gh workflow run nix.yml --ref BRANCH -f group=windows, or all four Intel Darwin outputs with gh workflow run nix.yml --ref BRANCH -f group=x86_64-darwin. The registered workflow can run on a branch containing this workflow file; it does not require that file on the default branch. Runs for the same workflow/ref queue without canceling an active run; manual groups have separate queues. Legacy PR build jobs also preserve running jobs; their cancellation policy for other events and release tags is unchanged. A two-minute planning job selects these existing groups; it does not evaluate or build Nix outputs. Each build job uploads its existing library ZIPs and CLIs under an artifact named for the build host, target group, and source SHA. Each native CLI also runs the encrypted-store and profile-password smoke test, which can be repeated with python3 scripts/nix/smoke-cli.py result-nix/BUILD_SYSTEM/exe:simplex-chat/bin/simplex-chat.

The Windows job first prepares only the named direct input outputs of the canonical wai-app-static component from the same Windows CLI project, using the normal 7,200-second per-derivation limit. It then builds that exact component with a 300-second limit; success is reused by the full Windows build. A failure keeps the job red and skips the full outputs. CI replays only that failed component in a separate diagnostic derivation, with ISERV_ARGS=-v and strace -f -tt -T -s 80 around compiler calls. It copies the exact selected Wine scripts and appends +seh,+loaddll to their pinned WINEDEBUG value, with an assertion that no other script bytes change. Only compile calls receive the copied -pgmi path. Compiler queries, source/plan, original arguments, DLL collection, Wine prefix and normal/profile selection stay unchanged. The trace has a 15-minute step bound and a 300-second per-derivation limit; strace is fetched from caches only. Logs and the selected derivation/input metadata are uploaded even on failure. Repeat the canonical gate on Linux with python3 scripts/nix/windows-th-gate.py prepare followed by python3 scripts/nix/windows-th-gate.py build; use trace only after a failed canonical attempt. These commands use the committed checkout and preserve the canonical Nix core setting. A fresh Wine prefix and full syscall tracing can consume most of the diagnostic budget; a timeout does not identify a deadlock. The diagnostic changes its derivation and process environment; it does not qualify a Windows artifact or change the failed result.

After a failed native x86_64 Linux library build, CI tries a separate derivation with -v2 -ddump-timings and GDB on compile calls and uploads native-linux-ghc-phase-trace.log, including when that trace fails. Compiler queries and the original source, plan, dependency units and GHC arguments stay unchanged. GDB follows the compiler's shell wrapper into the actual ELF program, preserving its injected package DB and -B arguments. On a stopped fault it records up to 80 frames per thread, registers, the PC symbol and disassembly, shared libraries and memory mappings. It does not dump the environment or a core file. Normal compiler exit codes propagate; faults and debugger errors fail the diagnostic. The canonical build remains failed. The trace runs one derivation at a time and inherits the canonical Nix core setting, preserving Setup's job count and the compiler job arguments. It has a two-hour wall-clock/per-derivation bound, including cache-only debugger setup; a debugger cache miss fails visibly without compiling GDB locally or remotely. The job's six-hour limit stays in place. A local --cores override also changes those job arguments and must be reported when comparing diagnostics. trace-native-ghc.nix accepts the original component via --arg original and an optional pinned GDB derivation via --arg debugger. Omitting debugger retains the phase-only trace. GDB leaves ASLR enabled; its host/runtime effects must be reported when comparing diagnostics.

When the pinned Intel Darwin ghc-boot component fails, CI captures a separate Setup diagnostic. It selects one exact failed ghc-boot derivation from the canonical log and checks it against the selected compiler project; a rebuilt compiler therefore does not leave a stale diagnostic path. Multiple matching failed components cause the probe to skip. It preserves the newly linked Setup executable's bytes and runs its actual configure arguments with an explicit capture mode. LLDB mode records bounded thread backtraces, registers, instructions and loaded images; compiler queries and other Setup commands pass through. Child failures remain failures, and the canonical job stays failed. The diagnostic has a two-hour step/per-derivation limit and one five-minute deadline for LLDB plus any fallback. Its log is uploaded as intel-setup-lldb-trace.log, including after failure. Only this probe disables the Nix sandbox to reach host tools; it records that runtime difference and the debugger version. Hosted Intel uses direct-abort mode: LLDB hosts the Python command without attempting debugger attachment, and the probe invokes Setup once with a diagnostic abort interposer built by the runner's existing Apple Clang. LLDB mode also retains a fallback for an explicit permission denial before a child PID is assigned. It reports up to 80 current-thread unwind frames, available callee-saved registers and image addresses. It symbolicates mapped register addresses without dereferencing them, then forwards to real abort. Its constructor restores the original loader environment before Setup starts compiler children and disables core output for the traced process tree. This adds a loaded image to Setup; it changes no debugger permissions or security settings. Nonzero exits and terminating signals remain failures. The hosted Intel replay captured 14 frames through evacuate1 and scavenge_loop1. A local Setup built with the same cached compiler and target dependency IDs reproduces the panic in 1.55 seconds, and its traced replay matches nine cloud frame offsets. Native ARM64 and Rosetta/x86 C controls also cover the helper. A successful probe is diagnostic evidence.

Hydra continues to consume the same hydraJobs attributes. The older application release workflow still uses its existing GHCup/Docker compiler settings; migrating that packaging and release path is separate from this Nix build lane. The legacy GHC 8.10 Linux job now reports container memory counters and kernel OOM records after a failed CLI build, with a one-minute diagnostic limit. Its observed SIGKILL alone does not establish an out-of-memory cause.

Hydra needs builders advertising aarch64-darwin and x86_64-darwin to run the corresponding Mac/iOS jobs. Check its builder pool when jobs remain queued. For cross builds, the pinned haskell.nix uses recursive-nix as a scheduling marker; these builds do not need recursive Nix operations. This checkout opts ghc914-sh Android ARM32 out of that restriction on Linux with patched QEMU 9.1–11, after passing normal target TH under emulation. Other target/compiler package sets retain the upstream guard, which also covers cross-compiler stalls without QEMU. Advertise the marker on a native x86_64 Linux builder for those jobs; keep it off Rosetta slots. Nix also requires the recursive-nix experimental feature when building derivations with this reserved marker. Enable both extra-experimental-features = recursive-nix and extra-system-features = recursive-nix on that builder. The GitHub workflow sets the complete experimental-features list, including nix-command and flakes, because its installer suppresses those defaults when this setting is present. It adds recursive-nix to both lists only for its native x86_64 Linux runner. The upstream fixes and regression jobs are in haskell.nix PR #2585.

To build both Android APK ABIs from the current checkout, preserving the same compiler and source, use the existing builder's source/skip-checkout flags:

scripts/android/build-android.sh -sg VERSION_CODE

This command also needs the Android SDK, Gradle, and its other existing tools. The APK reproduction script now checks out the supplied tag for both ABIs. Historical releases that used a separate -armv7a tag must be reproduced with their historical build scripts.

Migration validation

On 2026-10-16, all 21 existing Hydra derivations evaluated successfully at f26dbbc0d93a5c359427307b4c57f7f5e2123d69: 11 on x86_64 Linux, two on ARM64 Linux, and four on each Darwin architecture. This includes both Android ABIs, Windows, and all three static Linux targets. This establishes derivation evaluation; the completed artifact builds are listed below.

GitHub run 38939062012 tested head c5b7a04fe as integration merge 582e88ccf07d5da06cfb0e9d7e99d9133f2f56af. Both native Linux library/CLI groups and all four ARM64 Darwin outputs passed, including encrypted-store and profile-password CLI smoke checks. The production x86_64 Linux compiler also passed the ELF alignment regression. The x86_64 musl CLI linked with -fully-static, installed, and completed fixup. These qualify nine built outputs; the static group subsequently failed, and its artifact upload did not run.

The ARM64 Android normal and profiling interpreters compiled and linked with the API-aware symbol table. The next Android failure is SplitMix 0.1.3.2 calling getentropy at API26. The local fallback described above passes exact-source compilation and shared linking for both ABIs with native Clang21/LLD21 and the actual NDK headers/libraries. API35 objects remain byte-identical. The ARM32 interpreter table also compiles with its real arc4random_buf registration and remains byte-identical at API35. An earlier source variant passed actual NDK Clang18; the final Clang18 fixture and production package-tar evaluation were blocked before compilation by the local Linux VM/frontier. The final Android interpreter runtime and application builds still need CI.

The new dynamically C-linked i686 interpreter reaches its link step but fails because symbols.i386-musl.o references private musl __aio_close. Its real GHC arguments retain static Haskell libraries and omit -fully-static. This proves the link-policy change reached the component; the helper ELF and TH execution remain unqualified. Normal and profiling plans retain all 95 record identities: only three iserv configure arguments disable executable-static; the other 92 records are byte-identical. The changed project skeleton rebuilds auxiliary boot derivations. An i686/QEMU C control reproduces the old static-loader error and loads the same libz.so with dynamic C linkage.

The newer focused static run 38974180411 at 43262a5c0 built and installed both normal and profiling i686 interpreters after the private-symbol fix. Its x86_64 static CLI also built. The i686 wai-app-static Template Haskell splice then failed: the runtime linker rejected duplicate __x86.get_pc_thunk.bx and .ax definitions before unresolved lookups, a QEMU target SIGSEGV, and proxy EOF. The scoped COMDAT candidate cross-compiles both changed C files against real GHC 9.14 headers with the original warning messages unchanged. Its physical source producer changes exactly Linker.c, linker/Elf.c, and linker/Elf.h. The paired native x86_64 Linux runtime script passes with the exact existing GHC 9.14 compiler on the Hydra host: the original RTS rejects the COMDAT fixture, and the patched RTS preserves relocations in both load orders while rejecting ordinary strong and stale archive-owner conflicts. Both explicit RTS object pairs come from the same configured source; their link maps exclude the original archive objects. This isolated direct run completes in 4.3 seconds without a Nix build or configuration change. A real GCC15 -m32 PIC fixture then exposed zero-sized thunk signatures, which the initial sized-only candidate rejected. The corrected paired test passes in 4.8 seconds, including zero-size and mixed load orders, local-label preservation, and rejection of extra functions, data, global/weak definitions and partial extents. Extra definitions use distinct names in each object so unrelated duplicate symbols cannot satisfy negative controls. The Nix check and final i686 QEMU TH/application builds remain CI checks, including profiling execution.

Intel Darwin reproduced its Setup GC abort and produced the direct-abort trace used for the info-table diagnosis above. The local renderer regression passes, while the patched compiler bootstrap and full Intel outputs await CI. Windows is still building; its Basement follow-up passes the focused reproducer.

The native ARM64 Darwin library and CLI built with Nix at f26dbbc0d and passed python3 scripts/nix/smoke-cli.py /path/to/simplex-chat, covering encrypted stores and profile passwords. Their actual outputs contain ARM64 macOS objects and the expected library registration or CLI executable. Earlier native BBS checks also passed; the relevant native source, compiler, flags, and dependencies remain unchanged.

A native build with the iOS Swift/CommonCrypto flags passed mobile C API checks using the rolled archive produced by link-mobile-library.sh: Swift JSON, encrypted stores, UTF-8, reopen behavior, and wrong-key errors with copying and nonmoving GC. This ran on macOS before archive retagging. The CommonCrypto SQLCipher component also built with Nix and registered both Apple frameworks. The replacement RTS and text built with Nix and passed ABI, instruction-baseline, UTF-8, and GC checks.

Both full Nix ARM64 iOS library ZIPs, Swift JSON and tagged JSON, built in that GitHub run. The downloaded artifact matched its published SHA256, outer/inner ZIP CRCs, and all ten archive hashes recorded in CI. Each ZIP contained its five expected archive entries and 3,632 ARM64 Mach-O members with iOS platform markers. The earlier f26dbbc0d ZIPs contained 3,630 members and valid Hydra product markers. The production FFI archive previously passed an instruction-baseline check with unchanged inputs. The package uses the full Cabal v2 unit ID in its archive filenames; the existing Xcode project updater accepts those names. This verifies the packaged library, not execution after retagging.

The Android API26 ARMv7 ghc-internal built with all 292 Haskell modules and C sources. Its unit ID and ABI match the earlier build, and GNU iconv is retained in installed paths and consumer dependencies. Canonical base and the complete normal interpreter dependency closure also built under Rosetta.

On 2026-10-16, the normal ARM32 TH trial passed with the production compiler, interpreter, and all 40 target-library outputs. A target-side splice asserted System.Info.arch == "arm" and a 32-bit Int, then emitted 42. The resulting shared ELF32 ARM library exported the test function; a static C caller initialized the Haskell RTS and received 42 from the same compiled object. This inspects the shared library and executes a static caller, rather than loading the shared library on an Android device. The byte-identical upstream annotation test also compiled through the target interpreter.

The identical production interpreter reached Haskell main with both x86_64 QEMU under Rosetta and native ARM64 Linux QEMU. It has no PT_TLS segment and needed no TLS adjustment. The separate C caller retains the API26 CRT alignment fixture used by the C regression. The final QEMU version guard preserves the exact compiler, core, base, interpreter, and wrapper identities from this trial. These functional changes are integrated at 748becda3eac5173bdbf34ef4eec476a9215326f.

The trial used a normal-only diagnostic wrapper around the real production vanilla interpreter. The full mobile wrapper also references a profiling interpreter, so normal application builds can still compile profiling dependencies. Both normal and profiling ARMv7 ghc-internal archives subsequently built in the earlier b4a1cdd54 compiler context: each contains 283 ELF32 ARM objects, with profiling objects in the profiling archive. Their installed registration retains GNU iconv paths. This does not qualify the later native RTS compiler change or the profiling interpreter's final link and execution. Profiling interpreter execution, corrected Android/Windows builds, the complete static Linux group, Intel Darwin outputs, Android devices/APKs, and an Xcode application build remain unverified. Derivation evaluation and the normal trial do not establish that those builds and runtime checks pass.

Updating package hashes

1. Install nix, gawk and jq.

2. Start nix-shell from repo root:

nix-shell -p nix-prefetch-git

3. Run in nix shell:

gawk -f ./scripts/nix/update-sha256.awk cabal.project > ./scripts/nix/sha256map.nix