Files
msd-core/tests
sim 26384ca988 refactor(#4653): make validatePath module-internal
Phase 3 of epic #4636, stage 3a. ADR-4650 decision 2: the engine stops being a
public shape. The only exported containment surface is now assertWithinRoot /
tryWithinRoot / requireSafePath, none of which can hand a caller a usable path
when the answer is unsafe.

The src/security.cts diff is one keyword. The engine body is byte-identical —
the dangling-symlink existence-oracle closure, the ancestor canonicalization and
the separator-aware boundary test are untouched, which is the whole constraint
this phase operates under.

WHAT THE TRANSLATION COST, AND THE RULE THAT KEPT IT AT ZERO. Roughly thirty test
call sites consumed validatePath directly, including the two BLOCKER regressions
that are this refactor's safety net. Translating them all to
`tryWithinRoot(...) === null` would have looked correct and silently destroyed
one of them: BLOCKER-1 asserts the rejection reason contains "unresolvable
symbolic link", which is what distinguishes a DANGLING symlink from an ordinary
escape. tryWithinRoot returns a bare null and cannot tell those apart, so that
assertion would have degenerated into "it failed somehow" — and the
existence-oracle closure could regress with the test still green.

So the rule applied throughout is: an assertion on the rejection REASON goes
through assertWithinRoot, whose throw carries the engine's message verbatim; only
assertions on the boolean go through tryWithinRoot. Under that rule no coverage
is lost. BLOCKER-1 still pins "unresolvable symbolic link" and BLOCKER-2 still
pins the exact canonicalized resolved value.

Three success-path tests came out BETTER than they went in. They previously
carried `expected safe:true, got error: ${result.error}` as an assertion message;
routing them through assertWithinRoot means an engine regression now surfaces the
real reason in the failure itself rather than as a hand-built string.

The two describe blocks named after validatePath are renamed — a block named for
a symbol the module no longer exports is a false signpost.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-09-12 12:22:22 -04:00
..