fix(#4636,#4653): close the symlink hole, revert a wrong collapse, fix six review findings

The RED checkpoint and two orthogonal reviews found eight defects. All fixed here.

THE COLLAPSE THAT WAS WRONG — installer-migrations. Routing ensureInsideConfig's
containment decision through the realpath-based canonical predicate broke four
tests, and the failure message says it plainly: "migration path escapes
configDir: extensions/gsd.cjs". That module's entire contract is that a
symlinked managed path is snapshotted, restored and backed up AS A LINK and
never dereferenced. The canonical predicate dereferences, then rejects the
result for escaping configDir — so it destroys exactly the thing the module
exists to preserve. Reverted to lexical, with the ruling recorded above the
function so it is not collapsed a third time. normalizeRelPath is the real
pre-gate there; it throws on absolute paths and '..' before this check runs.

That makes THREE deliberately-retained implementations, not two, and they share
one shape worth naming: a realpath-based predicate is the wrong tool wherever a
symlink must be PRESERVED rather than resolved. CONTEXT.md and
docs/explanation/security-model.md are corrected — both previously described
ensureInsideConfig as collapsed.

THE MISSED CONSUMER. tests/security-prompt-injection.security.test.cjs
destructures validatePath from the compiled lib; un-exporting it turned five
tests into TypeError. It appeared in my own earlier search output and I did not
follow it up. Translated under the same rule as the rest: assertions on the
rejection REASON go through assertWithinRoot, boolean-only through
tryWithinRoot.

VALIDATE-ONE-PATH-USE-ANOTHER, FOUND TWICE MORE. This is the fourth and fifth
occurrence in this epic of the exact defect it exists to prevent.
  - scripts/check-glossary-refs.cjs decided containment on `token` and then
    stat'd a separately re-joined path.join(ROOT, token). The ContainedPath is
    now carried through to the probe, so the validated value is the probed one.
  - src/init.cts computed skillPathContained and DISCARDED it, re-joining from
    the raw input for the existsSync and read. The branded type exists to make
    that a type error and here it was inert.

AND THE OVER-CORRECTION OF THAT FIX, caught before it shipped. The first attempt
also substituted the validated value into the EMITTED `ref` for a global skill.
That value is a display token, not a path anything reads through — the only fs
access in that branch runs on the lexical path beforehand — so substituting it
changed emitted output two ways: it is realpath-resolved, so a symlinked global
skills directory would have emitted its resolved target instead of the user's
own path, and it came from path.join, so Windows would have emitted a backslash
where the template has a literal '/'. Restored, with the distinction recorded:
the containment check there is a GATE, not a path producer.

A TEST THAT COULD NOT FAIL. The first symlink regression planted its symlink
from inside a hooked fs.readdirSync and never asserted the planting happened —
if the hook did not fire, the "nothing was written outside" assertion passed
trivially, green against vulnerable code. It now asserts the plant, matching its
sibling. The other two were re-checked: one already asserted its equivalent, the
other plants synchronously and cannot silently no-op.

THE SYMLINK FIX ITSELF, now that the tests are proven red on the matrix.
isPathConfined is lexical by design and structurally cannot see a symlink; three
callers relied on it with no defense of their own. install-engine.cts:1608 and
install-profiles.cts:880 refuse to mkdir/write through a link — mkdirSync with
recursive:true does NOT throw on an existing symlink-to-directory, so a planted
link redirected the SKILL.md write outside the install root.
install-profiles.cts:755 refuses to read through one — statSync FOLLOWS links,
so an outside file's contents were returned and installed as a skill body. Each
mirrors the guard retired-artifact-cleanup.cts:77 already uses.

Severity stated accurately rather than dramatically: only the read at :755 needs
no race. _removeGsdEntries sweeps a pre-planted link at :1608 before the write
loop, and :880's stageDir is a fresh mkdtemp, so both of those require winning a
window. They are fixed as defense-in-depth, not as live exploits.

ALSO: the Changed changeset claimed "every command's observable behavior [is]
unchanged". Three rejection messages are reworded. It now says so, and says that
none of them reveals a host path it previously hid. A stale comment in
verify.cts still named validatePath; an init.cts warning hardcoded "resolves
outside the project directory" for a check that also rejects absolute paths, NUL
bytes and empty strings; and the rationale deleted with check-glossary-refs'
retired helper is restored, noting honestly that a rejected token is now
realpath-resolved before rejection rather than rejected by string comparison.

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

docs/explanation/security-model.md

@@ -162,17 +162,31 @@ module is the central security utility. It provides:
 - Shell argument validation: arguments passed to subshell commands are
   validated before use
 
-Two containment checks elsewhere in the tree are deliberately NOT routed through
+Three containment checks elsewhere in the tree are deliberately NOT routed through
 this predicate, because each is narrower or stricter rather than a second opinion.
+The common thread is that a realpath-based predicate is the wrong tool wherever a
+symlink must be *preserved* rather than resolved.
+
 The backup-restore gate in `gsd-core/bin/gsd-tools.cjs` rejects symlinks outright:
 the canonical predicate accepts a link whose target resolves inside the root, but
 for a restore that is still wrong, because writing through the link overwrites
 whatever it points at instead of materializing a regular file at the backed-up
-path. And `isPathConfined` in `src/external-descriptor-trust.cts` is lexical by
+path. `isPathConfined` in `src/external-descriptor-trust.cts` is lexical by
 design, because two install callers must validate a destination *before* the
-`mkdirSync` that creates it, where `realpath` cannot resolve. A lexical check
-cannot see a symlink, so callers that rely on it for a write-confinement
-guarantee must pair it with their own symlink refusal.
+`mkdirSync` that creates it, where `realpath` cannot resolve. And
+`ensureInsideConfig` in `src/installer-migrations.cts` is lexical because that
+module's contract is that a symlinked managed path is snapshotted, restored and
+backed up *as a link* and never dereferenced — resolving it would dereference
+precisely the links the module exists to preserve, and then reject them for
+escaping the config directory.
+
+A lexical check cannot see a symlink, so callers that rely on one for a
+write-confinement guarantee must pair it with their own symlink refusal. Three
+install call sites did not, and now do: a link planted at a capability skill's
+destination made `mkdirSync` succeed silently and redirected the write outside
+the install root, and a link planted at a capability's own `SKILL.md` was
+followed by `statSync`, so an outside file's contents were installed as a skill
+body.
 
 **Runtime hook: `gsd-prompt-guard.js`.** This hook fires on every Write or
 Edit call that targets `.planning/` files. It scans the content being written