Added DRR and WFQ golden fixtures and wire them into CI.

[baochunli]

Summary

Closes a CI coverage gap left after the TraceSpec port (#99): the lean-fixtures job builds
drr_check and wfq_check but never ran them on golden fixtures — only AQM and CUBIC had fixture
scripts. DRR/WFQ were exercised solely by the protocol-traces conformance matrix, which asserts
accept: true but not exact checker output or coverage. In particular, WFQ's class_count_ge_2
coverpoint — the seen accumulator introduced when WFQ moved onto TraceSpec in #99 — was not
golden-tested anywhere.

This PR gives DRR and WFQ the same fixture coverage CUBIC/AQM have, plus --coverage-out golden
fixtures, including a 2-class WFQ trace that directly locks down the class_count_ge_2 / seen
path. No Lean source changes; checker behavior is unchanged. Test infrastructure only.

What's added

lean/fixtures/drr/

• drr_single_class_accept — the real trace emitted by the DRR CI conformance run, used verbatim; plus a .coverage.expected golden.
• drr_deficit_conservation_reject — schedule row deficit_bytes 500→501 → REJECT: line 3: deficit_bytes mismatch: got 501, expected 500.
• drr_class_id_out_of_range_reject — enqueue row class_id 0→1 → REJECT: line 2: class_id out of range.

lean/fixtures/wfq/

• wfq_single_class_accept — the real emitted WFQ CI trace, verbatim; plus a .coverage.expected golden.
• wfq_two_class_accept — a 2-class trace (two flows, distinct class_id); its .coverage.expected contains class_count_ge_2, golden-testing the ported seen accumulator.
• wfq_finish_time_reject — enqueue row finish_time_ns 1000000→2000000 → REJECT: line 2: finish_time_ns mismatch: got 2000000, expected 1000000.
• wfq_depart_timing_reject — depart row departure_time_ns → 2000000 → REJECT: line 4: departure_time_ns must equal time_ns on depart.

Runner scripts (lean/scripts/): run-drr-fixtures.sh, run-wfq-fixtures.sh (clones of run-cubic-fixtures.sh), and run-drr-coverage-fixtures.sh, run-wfq-coverage-fixtures.sh (clones of run-aqm-coverage-fixtures.sh) — only the build target / binary / fixture dir differ.

CI (.github/workflows/leanguard.yml): four steps in the lean-fixtures job after the CUBIC step. The build step already builds drr_check/wfq_check, so no other change.

How the fixtures were produced

• Accept traces are captured verbatim from the CI conformance run's emitted drr_events.csv / wfq_events.csv — valid by construction.
• Reject traces corrupt a single constant-valued field of an accept trace, so the error message is deterministic (no float-recomputed "expected" value).
• The 2-class WFQ accept trace was simulator-generated (a two-flow / two-class WFQ config → wfq_check accepts by construction), avoiding hand-computing float finish/virtual times.
• Every .expected / .coverage.expected is a snapshot of an actual checker run, not hand-written (golden-master).

Verification

• All four scripts pass with true exit 0 (each fixture reports ok:).
• Negative control: tampering a golden makes the runner exit 1; restoring it returns to exit 0 — confirming the harness genuinely detects mismatches.
• wfq_two_class_accept.coverage.expected contains class_count_ge_2, proving the seen path is now golden-tested.

Notes

• Fixture CSVs are force-added (the repo's **/*.csv ignore rule also covers the existing cubic/aqm fixtures, which are likewise tracked).
• The 2-class WFQ generator config is intentionally not committed, matching the existing convention (AQM/CUBIC fixtures don't ship generator configs). Happy to add it under configs/testgen/ if preferred.

🤖 Generated with Claude Code