test(guard): the sibling check reads manifests; the lockfile disagrees

[noahgift]

check_workspace_siblings_pathed.sh proves no Cargo.toml DECLARES an in-tree
crate from crates.io. It passes, and its own header names the class it cannot
see: a dependency we legitimately take from the registry can drag an in-tree
name in TRANSITIVELY, and no manifest in this repo mentions it.

Currently resolved from crates.io despite being in-tree (7 names):

aprender 0.27.8   <- whisper-apr 0.2.8 <- {apr-cli, aprender-orchestrate,
trueno 0.17.5                              aprender-rag}
realizar 0.8.6
trueno-quant, renacer-core, provable-contracts-macros
batuta-common     <- bashrs <- aprender-compute-xtask (NOT optional)

aprender 0.27.8 is this monorepo depending on a published copy of ITSELF, 36
minors behind the workspace's 0.63.0 — the exact self-referential cycle APR-MONO
was meant to remove. #2471 closed 24 of 32 such collisions by fixing declarations
(lock 1493 -> 1402 packages); these 7 survive because they are transitive.

NOT implemented as cargo tree --duplicates, which is the instrument the
existing guard's header recommends and is structurally blind here. --duplicates
reports ONE PACKAGE resolved at two versions, but trueno, realizar and
batuta-common are [lib] names — the packages are aprender-compute,
aprender-serve, aprender-common. The lockfile holds exactly one trueno
package, the registry one, so it is not a duplicate and never will be. Measured:
0 of 8 caught under default features, 2 of 8 under --all-features, which no
workflow passes.

The lockfile is also the only feature-independent surface. Cargo.lock records
optional dependencies regardless of feature selection, which is why it sees the
six that cargo tree cannot reach without --all-features.

Severity note, against the audit that raised this as CRITICAL: 6 of 7 are
lockfile-only at HEAD. cargo tree -i aprender@0.27.8 exits 101 under default
features, so the tree does NOT currently compile two mutually-incompatible
truenos. Only batuta-common is actually built, via a non-optional bashrs in
aprender-compute-xtask. The live harms are the published-crate cycle and
lockfile/supply-chain pollution, not a broken build. Ratcheted rather than
failing outright, because closing them needs the transitive source cut (vendor
whisper-apr per APR-MONO, or drop its published aprender dep) and that is a
separate change.

Reuses the existing guard's own name extractor (workspace_sibling_names.awk),
so the two cannot disagree about what "in-tree" means — independently produces
the same 272 names the manifest guard prints.

Mutation-verified, each RED then restored GREEN:

• a new registry package colliding with an in-tree name -> "grew 7 -> 8, NEW: aprender-core"
• name extractor pointed at a missing dir -> vacuity RED (103 < 200)
• lockfile parser matching the wrong source prefix -> vacuity RED (0 < 500)

4-row case table covers the parsing traps: path-sourced packages excluded, git
sources are not registry sources, and a source line must not leak onto the
PRECEDING package block.

Refs #2481, #2471

Co-Authored-By: Claude Opus 5 noreply@anthropic.com