GeezerVison

2026.11.12.1: system facts as data, one arrangement per provider entr…

…y, a base package's single reading, runners as provider statements (#924)

mcpp 2026.11.12.1 (design record: .agents/docs/2026-11-12-ecosystem-architecture-optimization-design.md).

- A provider entry states one arrangement: `glibc-sysroot` serves aarch64-linux-gnu on linux-x86_64 from the base
  package xim:aarch64-linux-gnu-sysroot (revision 3); the request aarch64-linux is answered as before on every host.
- The link model is rendered, not inferred again; a base package has one reading; a dynamic artifact this machine
  does not load names the row's loader and carries no RUNPATH of this machine.
- System facts are data: [[platform.system]], provider `system-headers`, `tier` and `runner`, platform
  `runner.built-as`, row `loader` and `execution`; the engine's sources name fewer systems (ratchet in CI).
- A provider's runner is the lowest tier of the runner statement; its `requires` are installed on the first run.
- A program of another system with no runner is built, not run (exit 2), before anything is started.
- `[target.<sel>.abi] longjmp`.
- MSVC ABI with clang: the C line states the C library model and the C standard as two words.
- mcpplibs.mwrap.xz 5.8.3.1: `mcpp pack --compress xz` no longer faults on Windows.
- The openkal targets' platform is openkal-platform 0.27.1.
- Tests: a script that does not apply exits 77 and is a skip (mtest 0.6.2); the coverage check has `owed`.
  Six scripts that ran on no runner run again (13, 28, 31, 650, 653, 657).

Pull request: #924 (its body holds the self-review).

2026.11.10.2: FreeBSD's musl identity; cfg barewords from the rows; #918

; MSVC's C library is ucrt; the test fast path names the host (#920)

* The compiler witness asks again for a record Windows refuses; the matrix scan reports supervision lines
* test: a test of an unchanged project that this host cannot execute names the host's target
* prepare_inputs: the bareword vocabulary precedes the comment of unknown_tokens
* A platform refused for a row names the rows its providers serve
* Matrix: the FreeBSD musl rows and the 0.26.0 version group
* README: the FreeBSD musl rows are preview
* CHANGELOG and design record: barewords, #918, the chosen platform
* A platform chosen for a request serves it, by whichever selector chose it
* A namespace's package may be named version, path, git or workspace (#918)
* cfg barewords are the rows' systems, and a row states its family
* Design notes: K9 and the I1 correction in the implementation record
* 2026.11.10.2: FreeBSD's musl identity over openkal-platform 0.26.0
* The MSVC payload's C library is ucrt, as the MinGW payload's is

Co-authored-by: speak-agent <[email protected]>

2026.11.10.1: base packages, Windows GNU-ABI from every host, FreeBSD…

… rows, foreign packing, test records for any framework; mcpp#911 N1–N8, #913, #902, #917 (#916)

* E2, E13, E14: an entry per program's flags; the kernel-abi provider's interfaces for build programs; layer suppliers may not require one another in a circle

* E1: rows a platform defines are read as the built-in ones are; nothing of the payload's host side reaches a row whose C library a platform describes

* E3: a GNU-ABI Windows row finds its resource compiler in the paired payload or uses llvm-rc; a quoted toolchain header is the toolchain's

* E4: FreeBSD's identities are built-in data with their compiler triple; the rows' systems are words of the language; a platform-served row names the platform as its remedy

* E5: a base package states the C++ run-time and compiler run-time it holds beside its C library; its headers, libraries and std module stand where the payload's would

* E6: a patched member of a git repository is patched in a copy of the repository, so its edges to sibling members hold

* E7: a foreign artifact's closure is read from its DT_NEEDED; --mode static links and checks a static program

A program for another system was handed to the host's ldd (exit 127). Its
closure is now read from the file: a library this build produced travels
beside it, every other name is the target system's. `--mode static` sets
force_static and checks the file, so a dynamic program is never packed under
the `-static` name; a row whose C library a platform describes supports a
full static link.

* E8: [platform.link] flags that a compiler link does not use are named; docs/04 states per-target object variants

The words are given to ld.lld only; a platform with driver = "compiler" now
receives a warning that names [platform.build] ldflags. docs/04 states the
per-target object variants and per-program entry objects instead of the
retired compile-once sentence.

* E9: a MinGW-w64 base package serves the GNU-ABI Windows rows beside llvm on every host; the unserved-target remedy names a spelling the parser accepts

The mingw-llvm provider names xim:mingw-w64-sysroot (the target side of
llvm-mingw: MinGW-w64 on UCRT, libc++, libunwind, compiler-rt) for
x86_64-windows-gnu and the new aarch64-windows-gnu row, below the gcc payload
where that exists. A base package on a MinGW row is the driver's sysroot with
the run-times its layers state; its compiler runtime is read from the
directory the base names (compiler-runtime-dir), and a static link takes its
archives. A provider's sysroot package is the base statement of last resort.
The unserved-target diagnosis names [toolchain] <host> = { path, prefix }
instead of [target.X] toolchain = "<path>" (mcpp#913).

* E10: a scan override of a file this build excludes is kept for the builds that compile it (mcpp#902)

A [scan_overrides] glob whose files exist and take no part in this build -- a
feature that is off, a ! rule of a [target.<selector>.build] -- is no longer
refused; a glob that matches no file of the package still is.

* E11: the 32-bit hosted identities (i686-linux-musl, i686-linux-gnu, i686-openkal) are registered intent

Planned rows: a request is refused as planned rather than as an unknown
target, and an explicit [target.<row>] toolchain opts in early. No payload of
the index targets them, and openkal's 32-bit class is a design discussion.

* docs: SPEC-002 v1.6 (base packages, layer-cycle), SPEC-006 v0.11, SPEC-013 v1.7, SPEC-014 v2.2, SPEC-016 v1.1; MCPP_KERNEL_ABI_INTERFACES and layer-cycle in docs/30 and docs/50

* docs: base packages (docs/22), the targets every host reaches (docs/21), foreign ELF closure and the static check (docs/10), defined rows, served FreeBSD rows and link.flags (docs/35); en and zh

* 2026.11.10.1: version, CHANGELOG; CI bootstraps from 2026.11.9.1

* AGENTS.md: the repository's rules in their shortest form, linked to the documents that state them; design record 2026-11-10 (cross-platform ecosystem, fourth version)

* E5: a base package is an ordinary package -- provides answers its layers, [build] states them, the package root is the sysroot; a provider's sysroot is its identity's base

The [base] table (an architecture-keyed roots map, include-dirs, libs by link
form, compiler-runtime-dir) is replaced by the manifest vocabulary every layer
supplier uses: include_dirs, ldflags, std-module*, and one new [build] key,
compiler_runtime_dir (-resource-dir for clang, -B for gcc). One base package is
the sysroot of one identity; a platform's provider takes sysroot as a package
or { path }, ahead of the platform's own. The mingw-llvm provider names
xim:mingw-w64-sysroot@20261104-1.

* tests: the rows E4, E9 and E11 add -- capability pins, the windows set, the rows snapshot; freebsd is a platform name

* README: the rows this release adds; graph.cpp stays within its length limit

* A provider's base yields to a graph-supplied system; a carrier that names no hosts exists where the platform's runner does

CI on #916: a carrier for windows-aarch64 was built now that aarch64-windows-gnu
is a row, over the mingw-llvm base beneath openkal-windows (an x86_64
implementation). The provider's sysroot, an engine default, is no longer a
statement when the graph supplies the system, as the row's toolchain
convention already is not; and a carrier that names no hosts is built only
where the platform's runner exists (runner.hosts), so --systems all leaves the
others out with the reason.

* mcpp test states what its programs report as they report it: parts, case records, metrics and bounds; SPEC-017 (mcpp#917)

The runner's side of the test protocol, for any framework that implements it:
- test-parts: --cases names the parts [test] split runs a program as (parts),
  and each part's record names its program (part_of);
- test-case-stream: the report is read while the program runs (P5.1), and a
  case record states each case's beginning and end before the program's record;
- test-metrics: the cases' kind, tags, metrics and bounds and the program's
  usage reach the records, and the run kept under target/mtest carries them
  with each case's start and its failures' expected and actual values (P6,
  P10.2, P12), so a Bench case's measurements reach mtest analyze trend.
The --list main keeps its meaning. SPEC-017 states the runner, the protocol it
reads, the records and the store, framework-neutral.

* Windows resources: the subjects of the plan are collected by their own function

step13_windows_resources stays within the 400-line function limit.

* A provider's base is kept only once the graph is known; a paired sysroot is stated wherever no cfg is trusted

The base statement was memoised when first asked, before the dependency graph
was loaded, so a provider's base did not yield to a graph that supplies the C
library (x86_64-windows-gnu over openkal on hosts the mingw-llvm provider
serves). The compile side stated a paired provider's sysroot only when a cfg
was bypassed; the Windows llvm payload has no cfg, so aarch64-windows-gnu found
no libc++ headers there. The link side already stated it unconditionally.

* Target matrix: the cells of the mingw-llvm provider and the planned rows; e2e 1063 runs in the wine job

* Two mcpp processes of one home install one package once; e2e 182 states the MSVC ABI it asks for

A dependency is installed under the home's install turn (registry/.install.lock,
re-entered for the package's own dependencies); a process that waited for the
turn looks for the package again before installing it. A second process was
told the install succeeded while the first was still placing the package, and
stopped with "install path missing after fetch" (two e2e workers fetching one
package into one home).

e2e 182: llvm alone is not a request for the MSVC ABI now that clang pairs with
the MinGW-w64 payload and compiles against it on a host without a cfg; the
project writes x86_64-windows-msvc as its target.

* docs/25: a home's install turn

* A base package's C library is reported by the name its manifest states; the mingw-llvm provider names mingw-w64-sysroot 20261104-2 (mcpp:c-abi=ucrt)

The c-abi layer of a base package was reported by the package's name
(mingw-w64-sysroot); it is the C library its provides states. Revision 2 of the
MinGW-w64 base states ucrt, the name mcpp's vocabulary gives the Windows C
library and the one the MinGW-w64 payload answers, so that one requirement
(compat.glfw's Win32 backend) is met by either. The design records I1 in that
form, and the defects CI exposed.

* The base statement decided before the graph is kept as provisional, not dropped

base_statement_text points into the memo; an answer that was not kept left it
pointing at another target's statement or at none. The engine tier's answer is
kept and marked provisional, and asked again once the dependency graph is
loaded.

* A paired system's C++ library is stated on C++ commands only

-stdlib= rode the toolchain words every C and assembly command shares, so a C
unit of a paired target drew clang's unused-argument warning, an error under
-Werror. paired_stdlib_flag() is appended to the C++ flags and to the module
scan; the std module build asks for it as before.

* A payload not found after its install names the directories it was looked for in

The error "install path missing after fetch" now lists each version directory
the lookup tried, in its order, so that a provider that placed the payload
elsewhere is seen at once. e2e 15 prints the output of a failing mcpp test.

* A version a merge replaces withdraws the requests it made

A resolved record keeps its constraints by requester, and a merge that
replaces <ns>.<name>@<old> records that version as superseded: the items it
queued are not walked, and its constraints no longer bound a merge. The graph
os 0.4.4 -> 0.4.1 (pinned) was refused with "irreconcilable versions"
between openkal 0.24.0 and a request of openkal-linux 0.26.0, a version no
longer in it. e2e 1066 reproduces the order.

* mcpp test: the cases counted beside the programs, and the failures as an index

The summary states the cases that check (a case that declared cases is told
by them): cases: N passed, N failed, N skipped (in N programs). The failures
list each failed program and under it the failed leaf cases at the place of
their first failure; the per-program repetition of the case failures, which
the framework's own output already shows under the program's line, is gone.
The program count keeps its words.

* cenv: FreeBSD and NetBSD present POSIX as Linux does

The BSDs' ELF triples predefine __unix__ and are LP64 with a 32-bit wchar_t,
so a C library's posix/arch-default declaration is realised with no token
(<arch>-freebsd-musl: a C library above openkal-freebsd). docs/22 states the
row.

* The missing-payload diagnostic names the configuration its stores come from

* Design: tests and dependency resolution, a supplement (mcpp, mcpp test, mtest, mcpp-ide)

* Design: tests and dependency resolution, second revision (layers, the test session, case output and checks, self-review)

* Design: tests and dependency resolution, third revision (mcpp test spec with base and optional capabilities, the session as an optional capability, the global review)

* build: a package's root is its own file (package_root.cpp)

makePackageRoot and the expansion of include directories move out of graph.cpp, which had passed the 2,500-line limit; the graph's steps and the later phases take a root from there.

* config: a borrowed registry keeps the index places its owner recorded

A home whose [xlings] home names an existing registry wrote its own index directory and store into that registry's configuration. The e2e homes under a temporary directory did, and later builds over the shared registry were sent to a store deleted with its home ("install path missing after fetch" on Windows and macOS). Such a home now records no place of its own there; e2e 1067 fails before and passes after.

* e2e 1066: the replaced family's implementation is absent on every system

The test named openkal-linux; on Windows and macOS the pinned os brings the 0.24 family's implementation of that system instead.

* CHANGELOG: requests of a replaced version, the case summary, BSD C environments, borrowed registries

* test: release the build directory before running; build programs by the fast path; state only work done

- The build directory's lock is dropped once the tests are built, and each invocation reports into its own protocol directory, test-protocol/<pid>-<start>/, removed when it ends and swept when its process is gone (#915). e2e 1068 runs a build and a second test beside a running one; it fails before and passes after.
- The programs a test finds as MCPP_BIN_<NAME> are built by try_fast_build when no override applies, which now hands back the recorded programs.
- The packages are no longer announced before the build; Phase A's report names a package when one of its steps runs, and a test's Compiling line is written when the bulk pass ran a step of its own.
- The test graph is written to test.ninja (set on the three test plans).

* build: the test graph is test.ninja, and a scan both graphs name has one command

BuildPlan::graphFile names the file the backend writes and drives (-f). test.ninja shares the directory's objects, .ninja_log and .ninja_deps with build.ninja, so an output both graphs name with different commands is rebuilt rather than reused. The dyndep rule's text no longer depends on the plan: $module_map carries its own leading separator, and obj/main.cpp.ddi.dd was rescanned on every switch between mcpp build and mcpp test because its command differed by a space. The e2e that read the test graph from build.ninja read test.ninja.

* targets: the built-in platform of the openkal targets is openkal-platform 0.25.0

openkal-platform 0.25.0 (mcpp-index#751) groups openkal 0.25.0, openkal-image 0.5.5, openkal-musl 0.29.0 and the implementations of the 0.25 family; its mcpp floor is 2026.11.9.1.

* CHANGELOG: test.ninja, the programs' fast path, Compiling lines, the released build directory, platform 0.25.0

* test: the tests of an unchanged project run from the plan's record

A full mcpp test writes test-plan.json beside test.ninja (every test of the package: its program, its own sources, its protocol or the objects that state it, its guide; the runtime directories and the tool directory; [test] split and store) and a build-cache entry for the test graph (selection @test). The next mcpp test admits that entry through admit_recorded_build, which now takes the graph it replays, asks that the discovered tests are the recorded ones, that nothing under tests/ is newer than the graph and that ninja has nothing to do for the selected tests, and runs them from the record; the environment, the arguments after -- and the programs' places are joined at run time. A workspace's root package matches its own record (selection and group '.') on both fast paths. e2e 1069 fails before and passes after; on a workspace of 1,500 sources one program's unchanged tests went from 5.8 s to 0.13 s.

* e2e 1068: a test that builds its own project neither waits nor plans (#915)

The issue's shape: the test runs mcpp build in its project with the test's environment. Measured for T9 on 2026.11.9.1 and 2026.11.10.1, with gcc and with an llvm workspace: the test's environment (the subos's bin first on PATH, its library variables, MTEST_*) leaves the configuration's fingerprint as it is; the rebuilds the issue saw were the test graph written over build.ninja and the lock held through the run.

* test spec: SPEC-017 v2.0, launch records, framework capabilities and sessions, check details, a case's output

docs/specs/test.md is the mcpp test spec: SPEC-017 v1, the mtest protocol's normative text (P1-P12 kept) and the [test] keys, as base capabilities B1-B6 and optional ones O1-O9; test-runner.md is folded into it. B5: --no-run records carry each program's launch. B6: an entry target's test-capabilities and test-session, reported by --guide. O3: failure source and fatal, case checks, in records and the store. O8: a case's output from the program's output events, or from a program that ran that case alone; the output of several cases is not divided by arrival, which a buffered stream does not allow (measured). Features test-launch-records, test-frameworks, test-checks, test-output. e2e 1070 fails before and passes after.

* pack: the generated launchers depend on mcpplibs.os 0.4.4; e2e 1068 and 1069 on every host

The Windows launcher and the single-file launcher pinned mcpplibs.os 0.4.1, of the openkal 0.24 family, which the built-in platform 0.25.0 refuses (cross/mingw-cross, e2e 988); 0.4.4 ranges over the 0.24 and 0.25 families. e2e 1068 bounds its commands without timeout(1), which macOS lacks; e2e 1069 asserts that no dev-dependency is resolved, since the programs' build plans on hosts where mcpp build plans each time (macOS).

* dependencies: the root overrides a package for the whole graph (#914)

override = true on a root's dependency or in [workspace.dependencies] answers every request for the package in the graph its table governs ([dependencies]: build and tests; [dev-dependencies]: tests; [build-dependencies]: build programs). A replaced request that does not already admit the override is named once (Overrode); mcpp.lock marks the package and --locked reads the marker; an override of a platform version group's member is taken with a warning; a library's override is ignored and said to be. A refusal of two irreconcilable versions names the overriding line; a library's exact pin of a package every consumer shares (a target-side layer, a kernel ABI) is noted to its users and refused by mcpp publish. src/build/prepare/overrides.cpp holds the steps. e2e 1071 fails before and passes after.

* unit: the Windows launcher's manifest names mcpplibs.os 0.4.4

* manifest: test-capabilities and test-session are known target keys

An entry target that states them is not warned of an unsupported key.

* e2e 1070: the framework's path in host spelling

* link: the toolchain's C++ archives built for another target are not linked

flags.cppm reads the triple of the directory the payload's libc++.a was found in; another architecture or system marks MechanismInput::foreignCxxArchives, and resolve() links no C++ runtime there and states that the row's is a package's (mcpp:c++-abi). Measured on aarch64-freebsd-musl from a Linux host: ld.lld: libc++.a(algorithm.cpp.o) is incompatible. The OpenKal conformance suite now takes openkal-llvm-runtime there (openkal-freebsd).

* bench: mcpp's own costs as mtest Bench cases with a history (no-op build, why flags, one interface edit)

bench/tests/engine.cpp times the mcpp under test (MCPP) on a small module
fixture; each cost enters the mtest store as a metric with its samples. The
engine comparisons remain the standard set run by hand.

* CI: mtest pinned once and every test run reported through it

.github/versions.env pins mtest (MTEST_VERSION, MTEST_SHA256); the composite
action mtest installs it and exports MTEST, and e2e-mtest takes it from there.
test-report merges the stores mcpp test kept in a job, writes the job's
Markdown report into its summary and uploads the store; the unit jobs of the
three hosts and the Linux bench run use it. Each e2e shard writes its report,
run_all.sh groups the shards by the workflow run, and ci.yml's test-report job
renders one report over every job (mtest ci report).

* Design notes: implementation records of the ecosystem plan and its supplement

* CI: mtest is installed into the global subos, so the report runs it outside the repository

* resources: a cross gcc's Windows headers are the toolchain's; an include nothing provides is the script's own

- system_include_dirs reads a cross gcc's own <root>/<triple>-w64-mingw32/include when no payload is paired: under the mingw provider, a script's "winres.h" was taken for a file of the project that no action produces, and the build failed in ninja (e2e 1058, the cross gcc case).
- A quoted include found in no directory is an input only when a build action writes it; otherwise the script's preprocessing decides (#if __has_include) and the resource compiler names a missing one with the script and line (e2e 1072).
- docs/04 (en, zh) and CHANGELOG.

* Design notes: the resource findings of the validation in the implementation record

2026.11.9.1: targets and providers; a platform defines rows; mcpp#911…

… G1–G12, B1–B3; tests are mtest's (#912)

The target model of SPEC-016, the items of #911 as evaluated in the design's §19, a graph's C library leading a Mach-O image's link, and mtest 0.5.2.

Co-authored-by: speak-agent <[email protected]>

2026.11.8.1: mtest's test protocol in mcpp test, tests without main, …

…one std module builder per entry, lock targets per plan, identity-key manifests (#906)

mcpp test runs the runner's side of the mtest test protocol (version 1): a program lists, selects and reports its cases, nested selections included; a test file of mtest's objects needs no main; mcpp new scaffolds tests with mtest 0.4.0. Two builds that find a std module missing compile it once. A framework that is a dependency and a dev-dependency offers its beginning once. main is detected after the preprocessor. Lock targets per plan, manifests keyed by identity, the target budget, verified openkal rows. Docs, e2e (1032–1039) and the ecosystem design (W8–W10).

Co-authored-by: speak-agent <[email protected]>

2026.11.7.2: a launch description takes the fast path; the built-in p…

…latform is one release; an archived member test carries its files (#900)

Co-authored-by: speak-agent <[email protected]>

2026.11.7.1: compiler recovery, build generations, platform version g…

…roups and minimal header regions (#889)

Co-authored-by: speak-agent <[email protected]>

2026.11.5.1: scope build requests and report optional action progress (…

…#888)

Validate complete accelerator requests, bind capabilities in each consumer closure, isolate build-only target layers and add optional action progress with UTF-8 runtime channels. Align bilingual documentation and published platform defaults. The exact head passes full CI and the subsequent ecosystem review.

2026.11.3.1: qualify the openkal 0.23.3 release group (#884)

Release mcpp 2026.11.3.1 with built-in openkal-platform 0.23.3 and launcher os 0.2.9 over the fully published corrected Linux confinement graph. Apply platform flags to the std module and its cache identity, pass compression to package-format providers, reclaim runner caches with valid-version/native-path handling, and build every object of object-only packages. Preserve the extracted build-program function bodies. Align current English/Chinese documentation and record the qualified dependency group without presenting early warm-start samples as final resource measurements.

Validation: fresh native build, 164 root/toolchain-model programs, default public-group import-std image, E2E 01/942/951, version/documentation/module guards, full current-head PR CI, and the platform's complete four-host rerun pass. The Windows legacy-ACP helper positively exercises home and runner-cache paths. Published prerequisite source/index archives and both mirror hosts are byte-verified. The bootstrap pin remains on an existing installable compiler until this release is actually available.

Co-authored-by: speak-agent <[email protected]>

2026.11.2.2: openkal-platform 0.21.0; an embedded dependency's link r…

…equirements are its embedder's; response files as link inputs; cache fill (#883)

Co-authored-by: speak-agent <[email protected]>