Fix clol:test for regression learners, and migrate the test suite to rove

[masatoi]

Two pieces of work: a bug fix in src/rls.lisp, and a migration of the test suite from prove to rove.

1. clol:test was unusable for regression learners

clol:test always calls <TYPE>-TEST with both :quiet-p and :stream, but the -TEST
generated by define-regression-learner accepted only :quiet-p:

[UNKNOWN-KEYWORD-ARGUMENT] Unknown &KEY argument: :STREAM
  0: CL-ONLINE-LEARNING:RLS-TEST (src/rls.lisp:85)

So every (clol:test rls-learner data) signalled, and clol-predict could not be used with
RLS or SPARSE-RLS models at all. The classifier -TEST generated by define-learner has
accepted :stream all along — only the regression macro was left behind, which is why this
survived: regression had no test coverage.

The fix accepts :stream and writes one predicted value per test datum, mirroring the
classifier. The raw value is printed rather than its rounded sign, since that is the prediction
for a regression model. (when stream …) guards the write, avoiding the discarded-string
consing the classifier does when stream is nil.

End to end, after the fix:

$ ros roswell/clol-predict.ros t/dataset/a1a rls-a1a.model rls-pred.txt
Loading dataset
RMSE: 0.6596754
$ wc -l < rls-pred.txt
1605

2. prove → rove

t/cl-online-learning.lisp was one flat file of top-level prove is forms. It is now 36
deftest forms carrying 104 ok assertions — the same golden values, plus the new regression
coverage.

Why migrate. cl-mcp's run-tests tool reported ✓ PASS / Passed: 0, Failed: 0 for this
repository even when assertions failed. With one golden value deliberately corrupted, ros
reported × 1 of 96 tests failed while the tool reported PASS. Three layers compose into
that: the tool's :prove branch is an unimplemented stub that falls back to a runner treating
"no condition signalled" as success, and prove-asdf:run-test-system reports failure by return
value rather than by signalling. Filed upstream as cl-ai-project/cl-mcp#131. rove is the one
framework that tool supports end to end, including running a single test.

What changed.

• cl-online-learning-test.asd depends on rove instead of prove; :defsystem-depends-on (:prove-asdf) and the :test-file component type are gone. The component must be a
  plain (:file …) — rove resolves a system to its suites through component-source-files,
  which only walks asdf:cl-source-file, so :test-file would silently yield zero suites.
• test-op's :perform now fails on both ways a rove run can be green without being good:
  a red suite (rove reports failure by return value, so the usual idiom would leave
  asdf:test-system exiting 0) and a suite that registered no tests at all. See the CI section
  below for why the second check exists.
• Datasets load once at top level; each learner is built inside its own deftest. The prove
  file shared learners between the one-vs-rest and one-vs-one sections through top-level
  defvars rebound with setf; that coupling is gone, which is what makes single-test
  execution meaningful.
• The package shadowing-imports cl-online-learning:test — it name-conflicts with rove's
  re-exported rove/core/result:test under :use.
• CLAUDE.md updated. Worth knowing: the ros -Q run --eval … one-liner needs
  (sb-ext:disable-debugger) or SBCL enters the debugger when test-op signals, hits EOF on
  closed stdin, and exits 0 — a green exit for a red suite. ./t/run-test.ros is correct
  without that caveat and is now the documented default.

Golden values are unchanged. Every expected weight vector, bias and accuracy tuple was moved,
not retyped. Verified by extracting all 1771 numeric literals from the pre- and post-migration
files: the added multiset equals the deleted multiset, and the 96 original expected s-expressions
parse identically. Any value that had moved would have been a migration bug, not a value to
regenerate.

3. CI

Three changes, each prompted by something this branch's first CI run exposed.

• fail-fast: false. One broken leg was cancelling the other three before they reported,
  so two genuinely green SBCL runs and one unfinished CCL run all showed as failures.
• ./t/run-test.ros instead of rove cl-online-learning-test.asd. On ccl-bin the rove CLI
  printed not one assertion and still exited 0 — the same silently-green failure this
  migration exists to escape. Its run-file-tests resolves the .asd with
  (asdf:load-system (pathname-name file)) and no load-asd, and found no suites there. Going
  through test-op instead, CCL now runs all 104 assertions. The :perform guard described
  above would have caught the false green either way; it is kept as the structural backstop.
• ccl-bin × macOS-latest excluded. Those runners are arm64 and CCL no longer ships a
  Darwin binary Roswell can install (Not supported platform arm64), so the leg failed at the
  Roswell install step before any of this project's code ran.

Remaining matrix: sbcl-bin on ubuntu and macOS, ccl-bin on ubuntu — all three running the
full 104 assertions.

Testing

$ ./t/run-test.ros ; echo $?
...
✓ 1 test completed
Summary:
  All 1 test passed.
0

36 deftests, 104 assertions. The test-op guard was verified against all three outcomes: a
green suite exits 0, a corrupted golden value exits 1, and a suite with no registered deftests
exits 1 with No rove tests are registered for cl-online-learning-test.

New regression coverage — the library previously had none:

• regression-rls — RMSE, and that :stream emits one line per datum whose values equal
  rls-predict. A line-count-only check would pass even if the fix printed the wrong thing.
• regression-sparse-rls — the same for the sparse variant, which the same bug affected.

Known, deliberately out of scope

The regression -TEST uses (reduce #'+ (mapcar …)), so it rejects a vector test-data that
-TRAIN accepts via etypecase — the classifier -TEST uses count-if and handles both.
(clol:test rls-learner (coerce data 'vector)) signals a TYPE-ERROR. Empty test-data also
divides by zero. Both predate this branch; leaving them for a separate change kept this diff
focused on the keyword-argument bug.

A clean compile emits two style-warnings from src/cl-online-learning.lisp, where dim-of and
n-class-of reference the one-vs-rest / one-vs-one structure accessors before those
defstructs appear. Pre-existing; this branch does not touch that file.

🤖 Generated with Claude Code