- Strict child now gets the verified stdlib tree read-only at
/run/marqdo-stdlib with MARQDO_STDLIB_DIR set, so Compat-C 打印 works under
strict without exposing a host path (ADR-0021 update).
- Add a strict-lane negative: resource.open(file) is refused with
permission.denied; hello/factorial/multi-module and --json still pass.
- StrictHostCommand gains stdlib_dir; CLI discovers the bundle and sandbox
binds it; sandbox tests constructors updated.
Evidence: G-C lane PASS on the disposable Ubuntu 26.04 VM (7/7 adversarial
matrix, positive CLI, strict language-resource refusal, forged marker), plus
fmt/clippy/423 local tests.
WHY
- macOS/Windows failed to compile because `cfg!` left `linux::probe()` in the
type-checked branch (E0433).
- GitHub-hosted Linux can pass the strict probe and then fail the real launch
with EAGAIN; the CLI test treated the inspect snapshot as a permanent promise.
- The security matrix shared process-id-based scratch paths, so parallel
`cargo test --workspace` made its own tests clobber each other.
- Condition paragraphs silently ignored executable `Strong`/`Emph`, and
condition parse spans included the code-span backticks.
- The corpus had no native power/control-flow/inert fixtures and no independent
generic Tool consumers.
WHAT
- Dispatch strict detection through `#[cfg]` functions; add a non-Linux
compile/fail-closed regression test.
- Split `linux::run` into an injectable probe/spawn seam; test probe-OK ->
launch-EAGAIN returns LaunchFailed and never executes unrestricted.
- Make the Linux CLI contract accept only (a) successful isolated run or
(b) non-zero, diagnosed, fail-closed refusal with no project stdout.
- Give each security test a unique project/outside path and compile the helper
once.
- Reject executable roles in a condition paragraph; map condition expressions
from the literal bytes inside the backticks.
- Add 10 AST, 8 semantic, 4 compile-negative and 2 runtime-negative fixtures
for power, branch/loop conditions, source spans, inert prose/code fences and
error boundaries.
- Add `sample-view`/`sample-knowledge`/`sample-proof` independent Tool v1
packages and an end-to-end Analysis/Trace consumer test.
- Add a locked path-dependency content-tamper test.
TESTS
- cargo fmt --all -- --check
- cargo clippy --workspace --all-targets -- -D warnings
- cargo test --workspace --locked
- G-C VM lane (separate evidence commit/run)
CI_URL
- none from this host: the checkout has a Gitee remote and no GitHub
credentials. Windows/macOS CI remains UNVERIFIED.
BREAKING_CHANGE
- An executable node in a list-item condition paragraph is now a hard
`markdown.invalid_construct` error instead of being silently ignored.
SECURITY_SCOPE
- strict remains fail-closed; no downgrade, no new capability, L2 still
rejected. G-C still requires the dedicated self-hosted lane for CI-signed
evidence.
Strict mode now re-executes the CLI inside a Bubblewrap supervisor with
user/mount/PID/net/UTS/IPC namespaces, no-new-privs, all capabilities dropped,
a read-only /project, private /proc and /dev, rlimits, output capture and
wall-clock process-group cleanup. The child verifies PID 1, NoNewPrivs, empty
CapEff, an unshared uid map and loopback-only networking before project code.
- PlatformCapabilities::detect runs a real bwrap probe; no probe, no strict
- strict L1 supports pure Marqdo + Math; Foreign/ABI/worker/kernel/grants/trace
are rejected with exit 2 before sandboxing
- forged MARQDO_STRICT_ACTIVE is rejected; missing bwrap fails closed
- add adversarial helper, 7-test escape/resource matrix, seeded security lane
script and self-hosted security workflow
- executed G-C lane as uid 1000 in a disposable Ubuntu 26.04 VM: 7/7 PASS
plus positive CLI and refusal matrix; evidence in docs/evidence
- ADR-0021, spec/security.md, STATUS/COMPATIBILITY/README/CHANGELOG updated;
ADR-0019 remains for non-Linux/fail-closed rules