fix(deps): the monorepo pulled its own siblings from crates.io, and the gate that should have caught it could not fail

[noahgift]

fix(deps): the monorepo pulled its own siblings from crates.io, and the gate that should have caught it could not fail

CLAUDE.md states the rule: the workspace is a self-contained DAG and every sibling
must be a path alias consumed as { workspace = true }, because "pulling a
crates.io copy reintroduces a duplicate crate + the published-crate dependency
cycle the consolidation removed".

36 declarations violated it. trueno — the SIMD/GPU foundation — resolved at
0.16, 0.16.5 AND the in-tree 0.63.0 simultaneously. jugar-probar spanned seven
declared versions. All now resolve to the in-tree path.

The gate that should have caught this exited 0 by construction

Makefile:485 was:

@cargo tree --duplicate | grep -v "^$$" || echo "✅ No duplicate dependencies"
@cargo audit || echo "⚠️  Security issues found"

|| reads GREP's status, not cargo's. Measured on the unfixed tree: 1,828 lines
of duplicates printed, exit code 0. The second line swallows cargo audit
failures the same way. And nothing in .github/workflows/ or any tier target
invokes deps-validate at all, so even a working version was never run.

Third instance of this exact class here (#2336 qwen-story-daily, #2360 make
publish's POST-PUBLISH VERIFICATION). Both lines now redirect to a file and read
the real status; make deps-validate returns non-zero on the current tree, where
it returned 0 before.

The guard, and why it ships in the same commit

scripts/check_workspace_siblings_pathed.sh fails any manifest declaring a
workspace-local crate with a registry version =. It matches on BOTH [package] name and [lib] name, which is not optional here: package aprender-compute has
[lib] name = "trueno", so a checker comparing only package names is blind to the
single most important violation.

PASS: 100 manifest(s) scanned against 272 workspace-local crate name(s)
      (positive control flagged, so this absence is a measurement, not a blind spot)

self-test: 30 cases from a table file, plus 6 end-to-end tree probes.

Landing order was a real hazard. The guard is wired into guard-runner-labels,
which gate hard-requires with an explicit exit 1. Merged before the 36 manifest
fixes it would turn main RED and block every PR. Verified they pass together: with
both applied, the guard exits 0.

Numbers, and why the headline one understates it

cargo tree --workspace --duplicates: 57 -> 56 distinct duplicated names.

That is a poor measure of this change and should not be quoted as its value.
cargo tree --duplicates is structurally blind to the class being fixed: the
in-tree packages were RENAMED during consolidation (aprender-compute, aprender-serve),
so a registry trueno and a path aprender-compute are different names and no
collision exists for the duplicate detector to see. The verifier proved this —
cargo tree -p aprender-rag --features transcription lists crates.io
aprender v0.27.8 and trueno v0.17.5, a stale published copy of this monorepo,
while --duplicates on the same invocation reports nothing.

Cargo.lock is the honest measure: 1493 -> 1402 packages, 91 removed, 0 added.

Not in this commit

• The third-party version unification (57 -> 41) conflicts with this change on
  eight manifests. Hand-merging two sets of manifest edits is how a silent defect
  ships, so it rebases and lands separately.
• whisper-apr still transitively re-imports crates.io aprender 0.27.8,
  realizar 0.8.6, trueno 0.17.5 via apr-cli's whisper feature. Real, and
  invisible to --duplicates for the renaming reason above. Its own ticket.
• crates/aprender-test/Cargo.toml:102 declares probar = { path = "crates/probar" }
  pointing at a directory that does not exist. That crate is workspace-EXCLUDED
  and unbuildable already (APR-MONO consolidation debt: 126 verified findings — a member that will not compile, 9 orphans, guards scanning the wrong directory #2470).

cargo check --workspace: clean, 0 errors.

Refs #2463, #2470