* test(#3333): fold the runtime & install surface fix-* cluster — Wave 1
Folds 11 legacy tests/fix-*.test.cjs regression files into their module's
main test suite: 6 folded into existing suites (host-integration-descriptors,
effort-surface-axis, trae-imperative-reference, hermes-skills-migration,
gsd-agent-isolation-guard), 5 renamed to become the module's sole suite
(cursor-hook-workspace-roots, cursor-subagent-isolation,
lint-compiled-artifact-sync, hooks-commonjs-marker,
shared-hooks-dir-resolution). All 195 test() blocks preserved with zero
drops; lint-test-file-count.cjs and eslint remain clean. No production code
changed. Wave 1 of 7 in #3315 (H3 of epic #3053).
* test(#3333): replace try/finally with t.after() in isolation-guard tests
CONTRIBUTING.md bans try/finally inside test bodies (masks failures, not an
approved pattern). The fold in the prior commit carried 27 instances forward
verbatim from the deleted fix-3045-dispatch-isolation-resolver.test.cjs into
an otherwise-clean file. Converts each to the approved per-test t.after()
cleanup pattern — same cleanup call, registered instead of finally-wrapped.
No assertion, fixture, or test-name change; test( count unchanged at 50.
Found by the Standards review pass on Wave 1 (#3333, H3 of epic #3053).
* fix(#3333): restore raw NUL byte mangled by the fold in hermes-skills-migration.test.cjs
The prior fold commit copied fix-2284-hermes-agent-delegate-task-projection's
"collision-robust" test via a text-based Read/Write pipeline, which silently
turned a raw NUL byte (0x00) embedded in two string literals into a regular
space character. That corrupted the test's actual purpose (proving a NUL
byte survives a string-rewrite operation untouched) and produced a genuine
gsd-test failure: `24 !== 1` for `out.split(' ').length`, because splitting
on a space finds every space in the sentence instead of the single NUL byte
the test meant to isolate.
Root-caused by diffing the raw bytes (via `git cat-file blob` + `cat -v`)
between the pre-fold source and the folded target — confirmed exactly two
bytes differ. Restored via a byte-precise patch (latin1 round-trip) touching
only those two lines; test( count and every other byte unchanged.
* fix(#3333): use \x00 escape sequence instead of a raw NUL byte in test fixture
The prior commit restored a byte-exact raw NUL byte matching the original
fix-2284 source, and the production function (applyClaudeCodeBrandSwap) was
confirmed correct in a standalone repro. But the same raw byte still failed
through gsd-test's remote pipeline. Root cause is upstream of gsd-core: some
step in that transfer path does not carry a raw 0x00 byte through untouched.
A raw embedded NUL byte was never necessary here — `\x00` as a 4-character
escape sequence in the source text produces the identical runtime character
(U+0000) without ever putting a raw byte in the tracked file, sidestepping
any byte-oriented transfer step. Applied at both call sites (the fixture
string and the split() delimiter). No behavior change; test( count unchanged
at 76.
* fix(#3333): harden copyWithPathReplacement against a source file vanishing mid-copy (TOCTOU)
Surfaced by this PR's own gsd-test run: tests/install-minimal-hooks.test.cjs
and tests/opencode-command-dir-plural.test.cjs intermittently crashed with
ENOENT reading gsd-core/workflows/zzz-e5-drift-fixture.md. Root cause is
unrelated to test-file consolidation — tests/planning-prompt-drift.test.cjs
writes that fixture directly into the real, shared gsd-core/workflows/ tree
(main() hardcodes its scan root to the real repo) and deletes it in
t.after(); copyWithPathReplacement's readdirSync-then-read loop has no
protection against the listed file vanishing before it gets there, so a
concurrently-running install path can crash entirely on what is otherwise a
completely benign race.
Fixed by skipping (not crashing on) a listed entry that no longer exists by
the time the loop reaches it. Added a regression test that deterministically
reproduces the race (readdirSync snapshot still lists the file; it is
deleted immediately after) and proves both outcomes: no throw, and the
vanished entry's destination is never partially written.
Per CLAUDE.md's no-defer rule, a defect surfaced while verifying this PR is
fixed inline rather than deferred — this overrides one-concern-per-PR.
* fix(#3333): fix third NUL-byte-mangled occurrence missed by prior fix passes
The fold originally mangled three raw-NUL-byte occurrences to spaces, not
two — the earlier byte-restore and escape-sequence commits both only
targeted the fixture string and the split() delimiter, missing
out.includes('[ ]') a few lines below (should read out.includes('[\x00]')).
A remote gsd-test run kept failing on this exact assertion even after both
prior fixes, which is what surfaced the miss. Verified via a standalone
repro using the file's real (not retyped) fixture content: all six
assertions in the collision-robust test now pass. Zero raw NUL bytes remain
in the file; test( count unchanged at 76.
* chore(#3333): add changeset for the copyWithPathReplacement TOCTOU fix
Fixed-type fragment for the production defect fixed inline in this PR
(bin/install.js's copyWithPathReplacement). Exempt from docs/ requirements
per CONTRIBUTING.md (only Added/Changed/Deprecated/Removed require it).
* chore(#3333): backfill changeset PR number (pr:0 -> pr:3341)
---------
Co-authored-by: sim <sim@local>
33fca50d8a