[FEATURE] Make the Ruff lint and format checks actually gate CI, and clear the backlog behind them

[Nitjsefnie]

Is your feature request related to a problem?

Yes. The two Ruff steps in CI cannot fail the build, so nothing enforces them:

# .github/workflows/ci.yml:45-51
      - name: Run Ruff linter
        run: uv run ruff check . || true
        continue-on-error: true

      - name: Run Ruff formatter check
        run: uv run ruff format --check . || true
        continue-on-error: true

Both the shell-level || true and the step-level continue-on-error: true are present; either one alone is enough to swallow the exit code.

This looks deliberate rather than a slip, and I want to say so plainly: the MyPy step immediately below carries an explicit comment about the same construct (ci.yml:55-56 — "A missing target is a config error, not a finding: surface it instead of letting || true report success"), and the same shape is used for integration tests and coverage. So this is a question about intent and the debt that has accumulated behind it, not a report that something is broken.

Measured on main at 2ad932b, using the Ruff version uv.lock pins (0.15.22):

• uv run ruff check . → Found 228 errors, across 107 files (104 .py plus 3 notebooks under examples/)
• uv run ruff format --check . → 122 files would be reformatted, 153 files already formatted

Most frequent rule codes: TC003 28, UP006 13, E402 11, UP042 11, UP017 10, W293 9, B904 7. ([tool.ruff.lint] selects TCH; Ruff 0.15.x reports those findings under the renamed TC codes — same rules.)

The practical cost falls on contributors. CONTRIBUTING.md:84-96 tells them to run exactly these two commands, and both are red at HEAD, so someone new cannot separate their own violations from the pre-existing ones.

Describe the Solution You'd Like

Burn the backlog down, then remove || true and continue-on-error: true from the two Ruff steps so they actually gate.

Describe Alternatives You've Considered

• Leave as-is and document it. Cheapest — the steps stay informational. But the count keeps growing and CONTRIBUTING.md keeps pointing contributors at a command that is red before they touch anything.
• Gate formatting only, first. ruff format is mechanical and reviewable at a glance (122 files, no judgement calls), so it could be enforced well before ruff check is worked down.
• Ratchet instead of clearing. Baseline the current findings and fail only on new ones. Avoids a bulk diff entirely, at the cost of carrying the existing 228 indefinitely.

Use Case

The end state, for the linter step:

      - name: Run Ruff linter
        run: uv run ruff check .

      - name: Run Ruff formatter check
        run: uv run ruff format --check .

so that a PR introducing a new lint violation goes red instead of merging green.

Additional Context

Surfaced while working on #270, where an F401 in a new test file shipped green because of this — that one is fixed, but it prompted checking whether the gate could have caught it, and it could not.

Related: #250 (the MyPy step targeting a non-existent src/) was the same family — a gate that silently was not gating. The test -d openagent_eval guard added there is what makes the intent behind || true legible now.

If you would like the cleanup, I would suggest several reviewable PRs split by rule family rather than one 107-file diff, and the formatter pass kept separate from the linter pass since the two gates are independent.

Would you be willing to contribute this feature?

• Yes, I'd like to implement this feature
• I can help with testing
• No, I'm just suggesting