fix: 六处「答案已记录、做决定的代码不读它」+ 明确拒绝 system 工具链 (2026.8.30.2) - #538

…2026.8.30.2)

Analysis, measurements and design:
`.agents/docs/2026-08-30-issues-527-529-535-537-analysis-and-design.md`.

Two families. A record exists and the code that decides does not read it; or
the runtime search path mixes what was declared with what this machine happens
to have installed. The host-dependence policy that settles every severity here
is stated once in that document's §1.2 and in docs/03: mcpp and everything the
ecosystem publishes depend on no host, a user's own project may and that choice
is theirs to guarantee, and the boundary between a warning and a refusal is
whether the result builds and runs.

`mcpp test` re-derived an unchanged answer on every invocation (#529). Both
post-link ELF passes were written with a read-back keyed on the artifact's
stat, and `prepare_build` rewrites `resolution.json` from a fresh json object
at the start of every run, carrying neither record — so each invocation began
by deleting the memo its own backend was about to look for. The records move to
`.mcpp-runtime-verdicts.json`, which survives, and `resolution.json` keeps
publishing a copy, which is the shape `sync_resolution_verdict` already had.
Pruning now asks whether the artifact left the DISK rather than whether it left
the current command's plan: `build` and `test` share one output directory with
different link-unit sets and were deleting each other's verdicts. The record's
invalidation key gains the SubOS farm stamp and MCPP_ALLOW_HOST_LIBS, because
making a memo durable creates a staleness obligation that did not exist while
the answer was recomputed. Measured on ten link units, all warm: `mcpp test`
after a `mcpp build` 3.15s -> 0.36s.

A new source file in a `path` dependency was invisible to the fast path, which
swept only the project being built. `mcpp build` printed `Finished dev in
0.00s` and the module was never compiled. Content edits were caught, but by the
post-link snapshot rather than by the sweep. This is #359's shape in a
directory that fix did not reach, and workspace members depend on each other by
`path`, so it is not an edge case. The dependency source roots are recorded in
`.build_cache` and swept, manifest included.

`[toolchain] system` with a `build.mcpp` died as `posix_spawnp('') failed`
(#527 Bug 1). The resolved compiler path was in `tc->binaryPath` the whole time
and was not handed to the build.mcpp closure. This fills an unset variable; it
adds no host capability, and the same manifest without a build.mcpp already
built. The host-dependence warning states the cost once per build and names the
xim route.

`standard = 26` — the spelling #527 uses in three of its own examples — was
silently ignored, because the key is documented as a string and `get_string`
returns nothing for a bare integer. Both spellings are now accepted.

`[workspace.package]` and `[workspace.build]` (#527 Bug 2, RFC 3). The
workspace root's `[build]` reached no member. Scalars are inherited when the
member did not DECLARE the key, vectors append workspace-first, and
"declared" is recorded by both parse paths rather than inferred by comparing
against the default — a member pinning `standard = "c++23"` under a c++26
workspace must keep it, and that is the same bytes as the default. This is the
precondition the cpp20 design doc's §9-Q3 wrote down and deferred. Both
inheritance sites now call one function. `allow_host_libs` is refused there:
it turns a correctness gate off, and a root able to set it once would disable
it for members added later by someone who never read that file.

A dependency declaring a standard above the graph's is now reported instead of
silently discarded, degraded and promoted by --strict, and scoped to manifests
the author controls: every index descriptor with an mcpp segment declares
`language` (782 of 782 measured locally, 756 of 774 being C libraries carrying
a boilerplate "c++23"), so declaredness does not mean authorship there.

A dialect-class flag in `cxxflags` that never reaches the `import std` prebuild
is refused before compiling, naming `dialect_cxxflags`. Read from the effective
flag set, so a profile or target block is covered too; silent when nothing in
the graph imports std, and scoped to the root package because a dependency
carrying the flag may legitimately not import std at all.

Tests: tests/e2e/321..326, each with the negative case that keeps the check
honest. Docs: 03, 05, 06 in both languages.

Found by re-reading the merge, not by a failure. `[workspace.build]
include_dirs = ["shared/inc"]` was prepended to each member verbatim, so every
member resolved it against its OWN directory — looking for
`<member>/shared/inc` for a directory that lives at `<workspace>/shared/inc`.

This is #224 for a new key. `[indices].path` and `[workspace.dependencies]
path` are anchored to the workspace root for exactly this reason, and a third
relative-path key that skipped it fails as a missing header three members deep,
naming neither the manifest that declared it nor the root it was written
against. Anchored rather than refused: `expandIncludeDirs` already accepts an
absolute include directory, so the anchored form needs nothing downstream.

321 now includes a header from the workspace root, so the anchoring has an
assertion rather than a comment.