core/docs/analysis/full-suite-reds-ledger-2026-07.md
Shay dddd85cfca
All checks were successful
smoke / smoke (-m "not quarantine") (pull_request) Successful in 7m11s
lane-shas / verify pinned lane SHAs (pull_request) Successful in 25m7s
fix(tests): clear pre-existing full-suite reds (INV-21 pins + rotor_power)
Five fast-lane failures on main made the full suite dishonest.

- Refresh INV-21 allowlist pins in demo tests for holographic_vault (ADR-0241)
- Skip non-powerable (mixed-parity) rotor_power calls in generate recall and
  session anchor pull instead of crashing the cognitive path
- Align session_anchor_pull fixture to grade-1 field parity
- Ledger: docs/analysis/full-suite-reds-ledger-2026-07.md

[Verification]: targeted 10 passed; lane SHAs 9/9; smoke 175 passed
2026-07-15 15:39:09 -07:00

4.5 KiB
Raw Blame History

Full-suite reds ledger — 2026-07

Operator: Grok-4.5 (W2 pre-existing reds sweep)
Branch: fix/pre-existing-test-reds-v2
Base: forgejo/main @ 54228d45

Enumeration

Lane Command Result
Fast pytest -m "not quarantine and not slow" -n auto -q 5 failed, 11660 passed, 106 skipped (pre-fix)
Fast recheck same after fix (see verification below)
Slow pytest -m "slow and not quarantine" -n auto -q running / recorded at PR time
Full pytest -m "not quarantine" union of fast + slow

The briefs “~31 reds” appears to reflect an older main tip. On current forgejo/main, the non-quarantine fast lane surfaces exactly 5 dedicated failures. Smoke/pack/lane gates stay green because they never execute these nodes.

Red list and disposition

# Test Class Cause Disposition
1 tests/test_deductive_entailment_authority_demo.py::test_inv21_and_inv29_allowlists_are_unchanged (b) pin drift Demo pin omitted ADR-0241 writer core/physics/holographic_vault.py already on canonical INV-21 allowlist Fixed test — pin updated to match ALLOWED_VAULT_WRITERS
2 tests/test_proof_carrying_promotion_demo.py::test_inv21_and_inv29_allowlists_are_unchanged (b) pin drift Same as #1 Fixed test — same pin update
3 tests/test_engine_loop_proof.py::test_generated_final_state_satisfies_versor_condition_by_construction (b) producer _recall_state called rotor_power on mixed-parity transitions (multi-token even field ↔ grade-1 vault hit → odd unit versor). Exact rotor_power correctly refuses non-rotors Fixed producergenerate/stream.py skips non-powerable weighted hits (try/except ValueError)
4 tests/test_register_pack_convivial_v1.py::test_runtime_chat_turn_idx_distinct_under_convivial (b) producer Session anchor pull called rotor_power on non-rotor transition during finalize_turn Fixed producersession/context.py skips pull when power fails
5 tests/test_session_coherence.py::test_session_anchor_pull_output_satisfies_versor_condition (b) fixture + producer Test used even-grade _random_rotor against grade-1 anchor (mixed parity); also needs producer guard Fixed test fixture to grade-1 reflector + producer guard in #4

Classification key (from brief)

  • (a) stale committed artifact → regenerate via sanctioned generator
  • (b) telemetry/schema/pin drift → fix test if wrong, else producer
  • (c) genuinely broken capability → quarantine + ledger owner-thread

No (a) artifact regens and no (c) quarantines in this sweep. QUARANTINE remains empty.

Root-cause notes

INV-21 allowlist pin drift

Canonical allowlist lives in tests/test_architectural_invariants.py and already includes holographic vault (ADR-0241). Two demo “unchanged” pins were stale copies.

Mixed-parity rotor_power calls

  • Multi-token inject yields even-grade field states.
  • Single-token vault entries are often grade-1.
  • word_transition_rotor(A,B) = B * reverse(A) between different parities is a unit odd versor, not a rotor.
  • Historical rotor_power returned identity for non-simple cases (silent no-op). Exact power fail-closes; callers must not crash the cognitive path.
  • Guard matches existing word_transition_rotor ValueError continue/return pattern. No assertion weakened; algebra fail-closed kept.

Owner-thread follow-ups (non-blocking)

Thread Why
Grade-parity policy for field vs vault entries Soft-weighted recall currently skips mixed-parity hits; a future design may map grade-1 hits into even motors or store multi-token finals consistently
Demo allowlist pins Prefer importing/asserting equality against the single canonical frozenset to prevent re-drift

Hard-rule checklist

  • Never weakened an assertion to pass
  • Never deleted a test
  • No wrong=0 surface touched
  • No lane-pin artifact regen (N/A)
  • Did not edit .github/workflows/, core/physics/chiral_gate.py, or core/physics/goldtether.py

Verification (record at ship)

[Verification]:
  targeted fixed reds + modules: 10 passed
  python scripts/verify_lane_shas.py: 9/9 match pinned SHAs
  uv run core test --suite smoke -q: 175 passed
  pre-fix fast lane enumerate: 5 failed / 11660 passed / 106 skipped
  post-fix: those 5 green; no quarantine entries added