fix(scripts): give harness self-test an independent tally

[kyle-sexton]

Closes #3283

Summary

scripts/lib/test-harness.test.sh sources the library it tests, so a break in test_harness::report's return path or in fail() counter arithmetic is detected by the suite's assertions and then cannot be reported to CI. The process exit status is produced by the code under test.

Two mutations against a throwaway copy of the harness demonstrate it: deleting return 1 from report, and removing the increment from fail(). Both print FAIL: lines and still exited 0. Checking _test_harness_fail after report catches the first mutation but not the second.

Fix

• Parent assertions go through self_ok/self_fail, which keep an independent pass/fail tally the harness functions do not write.
• After test_harness::report, a non-zero independent tally forces exit 1 even if the harness counters or return were sabotaged.
• The last-line discipline still requires every other sourcer to end on test_harness::report; this self-test is the one file allowed to follow the call.
• A throwaway-copy mutation regression runs the self-test against both defects and asserts it still exits non-zero.

Verification

• bash scripts/lib/test-harness.test.sh — PASS=18 FAIL=0 (includes both mutation cases)
• bash scripts/check-silent-revert.test.sh — PASS=130 FAIL=0 (installs an EXIT trap and sources the harness)
• bash scripts/check-skill-portability.test.sh — PASS=92 FAIL=0 (installs an EXIT trap and sources the harness)
• shellcheck --rcfile=.shellcheckrc -x and shfmt -d clean on scripts/lib/test-harness.test.sh
• bash -n clean on scripts/lib/test-harness.test.sh

Related

• Refs scripts/lib/test-harness.sh: self-test omits the EXIT-trap interaction its design rests on, and re-sourcing silently zeroes the counters #3268 — EXIT-trap and re-source gaps that cannot report a sabotaged report/counter
• Refs fix(scripts): preserve test-harness counters on re-source #3271 — those cases landed; they detect both mutations and still cannot report them
• Refs refactor(scripts): unify test-harness scaffolding in scripts/lib/test-harness.sh #3248 — the shared harness this self-test holds to its exit contract

[claude]

Claude finished @kyle-sexton's task in 42s —— View job

---

Security review — no findings

• Gather context (PR diff, changed files)
• Run skip gate check
• Review changed file for security-relevant issues
• Post findings / conclusion

Scope: scripts/lib/test-harness.test.sh (+126/-34), the only file changed in this PR.

Skip gate: the change has no security-relevant surface. It's a local, self-contained test file: no network calls, no secrets/credentials, no user-controlled input. The new mutation-regression helper (run_mutated_self_test) copies fixed local files (test-harness.sh, the self-test itself) into a mktemp -d scratch directory and applies one of two hardcoded grep -v/grep -Fv mutations selected by a literal case value (drop-return-1 / drop-fail-increment) — not by any external or attacker-influenced input. mktemp -d ".../harness-mut.XXXXXX" gets a securely randomized suffix, so there's no predictable-tempdir/symlink race, and the scratch dir is removed after each run. No eval, no sourcing of untrusted content, no injection surface into the constructed shell commands.

No security issues found in this diff.

[github-actions]

Last security-reviewed head: f5a96eb777f38eeca9059ae882845439f7411e98. On the next push, the relevance gate compares only the commits since this SHA; delete this comment to force a full re-review.

[claude]

Claude finished @kyle-sexton's task in 27s —— View job

---

/review:code-review — skipped

• Gather context
• Check skill skip gate

This PR is already merged, which trips the /review:code-review skip gate ("PR is closed or not open"). Per that skill's instructions, review stops here with no findings posted — merged code isn't an active review surface.

If a fresh review is wanted (e.g. against a follow-up branch), please re-trigger @claude on an open PR.

[github-actions]

Claude has reviewed this PR 1 time. The lane skips further automatic reviews after 5; deleting this comment resets the count.