ci: drop continue-on-error on test-run step (bisect_ppx instrumentation)

The "Run tests with bisect_ppx instrumentation" step in
`.github/workflows/ci.yml` carried `continue-on-error: true`. Result:
test failures (including QTT typechecker test regressions) reported
as a CI pass.

The intention was apparently to keep coverage-report generation
working even when tests fail, but the better split is:
- TESTS must hard-fail on regression (this commit).
- Coverage-report generation may soft-fail (kept as `continue-on-error`
  on the "Generate HTML coverage report" step, since coverage tooling
  artefact issues should not mask test results).

This is the same anti-pattern class as echo-types' N5Falsifier xfail
gate (resolved via echo-types#136) and echidna's Tier-2/3 live-test
soft-fail (resolved via echidna#118).

The other 4 `continue-on-error` instances in this file are preserved:
- `Lint with odoc` (line 105) — docs lint, not proof validity.
- `Build bench targets` + `Run benches` (lines 127, 131) — benchmarks.
- `Generate HTML coverage report` (line 183) — coverage tooling
  artefact, not proof validity.

Refs hygiene audit 2026-05-28
(`standards/adoption-readiness-grades/AUDIT-FINDINGS-2026-05-28.adoc`).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: hyperpolymath

.github/workflows/ci.yml

@@ -175,7 +175,6 @@ jobs:
         run: opam install . --deps-only --with-test --with-doc
 
       - name: Run tests with bisect_ppx instrumentation
-        continue-on-error: true
         run: |
           opam exec -- dune runtest --force --instrument-with bisect_ppx