Adversarial Singularity Defect Ledger

Version: 1.6 Date: 2026-05-21

---

Purpose

This is a living document. It records every defect and test-design finding produced by the adversarial singularity test suite. Each entry states the test that finds it, what the engine does, and what correct behaviour looks like. When a defect is fixed, the entry is moved to the Resolved section with the fix commit and the corrected test result.

The suite does not validate correctness. It attacks mathematical seams, boundary conditions, and degenerate inputs to find where the engine lies or collapses silently. A test that fails on first run has found a real defect. A test that passes has proven the engine handles that singularity.

---

Test Files

Run command:

python -m pytest tests/unit/test_adversarial_singularities.py tests/unit/test_adversarial_house_singularities.py tests/unit/test_adversarial_dual_singularities.py tests/unit/test_adversarial_triple_singularities.py tests/unit/test_adversarial_hardmode_singularities.py tests/unit/test_adversarial_quad_singularities.py tests/integration/test_oracle_hostile_semantic_audit.py tests/integration/test_oracle_hostile_semantic_audit_wave2.py -q

Campaign result as of 2026-05-21 (current main):

• Pure adversarial corpus: 162 collected - 162 pass, 0 fail, 0 skip
• Adversarial + oracle-hostile campaign: 203 collected - 203 pass, 0 fail, 0 skip

---

Resolved Defects

DEF-001 - ecliptic_to_equatorial does not normalise RA output at 360°

• Test: test_layer1c_360_degree_input_normalises_to_zero
• Layer: 1c
• What happened: ecliptic_to_equatorial(360.0, 0.0, obliquity) returned RA = 360.0° instead of 0.0°. Input 360° is mathematically 0°, but the atan2 * RAD2DEG % 360 expression produced a tiny-negative intermediate value whose IEEE 754 addition to 360.0 did not underflow the representable floor, yielding 360.0.
• Fix: Changed math.atan2(y, x) * RAD2DEG % 360.0 to normalize_degrees(math.atan2(y, x) * RAD2DEG) in ecliptic_to_equatorial.
• Fix commit: see PR fix/coordinate-normalize-boundary
• Verified: 2026-05-21 - test passes; killer suite 42/42.

---

DEF-002 - icrf_to_ecliptic accepts zero-vector input silently

• Test: test_layer1f_icrf_to_ecliptic_zero_vector
• Layer: 1f
• What happened: icrf_to_ecliptic((0.0, 0.0, 0.0), obliquity) returned (0.0, 0.0, 0.0) - distance 0, but finite longitude and latitude - with no exception raised.
• Fix: Added if distance == 0.0: raise ValueError(...) immediately after computing the magnitude in icrf_to_ecliptic. Also applied normalize_degrees to the longitude output for consistency.
• Fix commit: see PR fix/coordinate-normalize-boundary
• Verified: 2026-05-21 - test passes; killer suite 42/42.

---

DEF-004 - Year-zero calendar round-trip broken

• Tests: test_layer3b_year_zero_calendar_round_trip
• Layer: 3b
• What happened: calendar_from_jd(julian_day(0, 1, 1)) returned day = 3 and year = 0 instead of day = 1. calendar_from_jd used the Julian calendar path (A = Z) for JD < 2299161, while julian_day uses the proleptic Gregorian formula (B = 2 - A + floor(A/4)) for all epochs. The two algorithms were inconsistent for pre-reform dates.
• Fix: Removed the if Z < 2299161: A = Z branch from calendar_from_jd. Now both functions use the proleptic Gregorian formula for all epochs, making julian_day(y, m, d) -> calendar_from_jd a round-trip for any epoch.
• Fix commit: 2026-05-21, moira/julian.py
• Verified: 2026-05-21 - test passes; adversarial suite 122/122.

---

DEF-005 - Out-of-coverage JD raises KeyError instead of OutOfRangeError

• Test: test_boundary_ownership_out_of_coverage_raises_not_silence (JD = -4 000 000)
• Layer: cross-cutting
• What happened: KernelPool.position() raised KeyError("No kernel in pool covers ...") when no segment covered the requested JD. DE441's minimum coverage is JD ≈ -3,100,015 (≈ 8500 BCE); JD = -4,000,000 falls outside this range.
• Fix: Replaced the raise KeyError(...) path in KernelPool.position() with raise OutOfRangeError(..., out_of_range_times=True).
• Fix commit: 2026-05-21, moira/spk_reader.py
• Verified: 2026-05-21 - test passes; adversarial suite 122/122.
• Related TDF: TDF-003 - JD = 0.0 and JD = -1,000,000 ARE within DE441 coverage and correctly return positions; see TDF section.

---

DEF-006 - Deep historical calendar round-trip error of 26-38 days

• Tests:
  • test_layer3g_deep_historical_calendar_round_trip[1.0]
  • test_layer3g_deep_historical_calendar_round_trip[500000.0]
  • test_layer3g_deep_historical_calendar_round_trip[-100000.0]
• Layer: 3g
• What happened: calendar_from_jd -> julian_day round-trips at deep historical JDs returned errors of 26-38 days. Root cause is the same as DEF-004: the Julian calendar path in calendar_from_jd was inconsistent with the proleptic Gregorian path in julian_day.
• Fix: Same as DEF-004 - proleptic Gregorian for all epochs in calendar_from_jd. All three parametrized cases now round-trip exactly.
• Fix commit: 2026-05-21, moira/julian.py
• Verified: 2026-05-21 - all three tests pass; adversarial suite 122/122.

---

DEF-003 - normalize_degrees(-1e-15) returns 360.0

• Test: test_layer1h_negative_epsilon_normalises_near_zero
• Layer: 1h
• What happened: normalize_degrees(-1e-15) returned 360.0. (-1e-15) % 360.0 computes 360.0 - 1e-15, but 1e-15 is smaller than the ULP of 360.0 (~5.7e-14 in float64), so the subtraction is a no-op and the result is 360.0.
• Fix: Added a clamp after the modulo: return 0.0 if result >= 360.0 else result.
• Fix commit: see PR fix/coordinate-normalize-boundary
• Verified: 2026-05-21 - test passes; killer suite 42/42.

---

Open Defects

(All defects resolved as of 2026-05-21.)

---

Compound Coverage

The original adversarial singularity suite proved that Moira survives isolated seams. The compound suites extend that proof to seam composition: two lawful boundaries under simultaneous pressure (dual) and three lawful boundaries under simultaneous pressure (triple).

Current compound corpus as of 2026-05-21:

• Dual suite: 12 tests pass. Covers segment boundary x route equivalence, polar fallback x Aries-axis normalization, polar fallback x public chart equivalence, deep historical calendar regime x chart product, and exact cusp equality x house ownership.
• Triple suite: 9 tests pass. Covers critical latitude x cardinal MC x public fallback chart, coverage edge x TT/UT conversion x public planetary position, body-on-cusp x polar fallback x exact equality rule, and segment boundary x TT/UT x equatorial route continuity.
• Hard-mode suite: 10 tests pass. Covers topocentric DE441 seam continuity on the public sky-position path, chart admission and rejection near coverage edges, polar fallback plus downstream lots and angularity parity, pole-adjacent subnormal-vector doctrine, and station-neighborhood truth across nearby JD boundaries.
• Quad suite: 7 tests pass. Covers segment boundary x apparent/equatorial/horizontal route coherence, coverage edge x TT/UT x chart assembly x topocentric public path, critical latitude x fallback x exact cusp equality x lots parity, and station x wrap x JD boundary x retrograde truth.

Compound-suite doctrine:

• acceptable outcome: finite canonical result or a named exception
• forbidden outcome: silent semantic drift
• public-path preference: attack chart, house, and planetary products before internal helpers when both are available

---

Oracle-Hostile Coverage

The first adversarial waves proved seam survival. The oracle-hostile waves ask a harder question: can Moira remain finite, smooth, and internally coherent while still telling a plausible lie against stronger cached authority?

Current oracle-hostile corpus as of 2026-05-21:

• Wave 1: 6 tests pass. Covers public houses(datetime, ...) against selected cached Swiss rows, public sky_position() plus Gauquelin downstream semantics against cached Swiss sector rows, and public find_phenomena() / get_transit() against cached Horizons rise/set rows with altitude-crossing and meridian-identity checks at the solved epochs.
• Wave 2: 35 tests pass. Covers public fixed_star(datetime) against cached Swiss fixed-star rows, public fixed-star propagation against ERFA anchor-star cases across centuries, and public twilight(datetime, lat, lon) against published USNO twilight tables with explicit solar-altitude-crossing semantics.

Oracle-hostile doctrine:

• acceptable outcome: agreement with cached authority within the declared route-specific budget
• forbidden outcome: a finite, normalized, persuasive result that fails its semantic contract against stronger authority
• route-specific budgets are explicit when public datetime wrappers lawfully pass through UT1/TT conversion before reaching the substrate engine

---

Test-Design Findings

These are not engine defects. They are historical test-design findings in which the suite's own hardcoded constants or coverage assumptions were too weak to support the intended assertion. All three have been corrected in the current suite.

---

TDF-001 - Station R/D JD constants were off by 0.5-1.0 days

• Tests: test_layer2c_retrograde_station_speed_sign_change[Mercury/Venus/Mars]
• Layer: 2c
• What happened: All six station R/D constants were rounded to noon (0.0 or 0.5 JD) and were off by 0.5-1.1 days from the actual zero-crossing. At +/-1 hour, the speed sign had not yet changed.
• Resolution: Replaced with bisect-precise constants (sub-2-minute accuracy):
  • _JD_MERCURY_STATION_R_2023 = 2460055.853
  • _JD_MERCURY_STATION_D_2023 = 2460079.633
  • _JD_VENUS_STATION_R_2023 = 2460148.562
  • _JD_VENUS_STATION_D_2023 = 2460191.554
  • _JD_MARS_STATION_R_2022 = 2459883.051
  • _JD_MARS_STATION_D_2023 = 2459957.370
• Resolved: 2026-05-21

---

TDF-002 - Moon perigee JD was 0.29 days too early

• Test: test_layer2j_moon_distance_local_minimum_near_perigee
• Layer: 2j
• What happened: _JD_MOON_PERIGEE_2023 = 2459966.08 was off by ~0.29 days from the actual distance minimum. The golden-section minimum search placed the actual perigee at JD = 2459966.369.
• Resolution: Updated constant to 2459966.369 (golden-section precise).
• Resolved: 2026-05-21

---

TDF-003 - JD = 0.0 and JD = -1,000,000 are within DE441 coverage

• Tests:
  • test_layer3c_jd_zero_raises_named_exception (renamed ..._returns_finite_position)
  • test_layer3c_deeply_negative_jd_raises_named_exception (renamed ..._returns_finite_position)
• Layer: 3c
• What happened: Tests were written expecting both JDs to raise OutOfRangeError, under the assumption that they fall outside DE441's coverage. Measurement showed DE441's minimum JD is ≈ -3,100,015 (≈ 8500 BCE). JD = 0.0 (≈ 4713 BCE) and JD = -1,000,000 (≈ 7451 BCE) are both inside the kernel range; the engine correctly returns finite positions.
• Resolution: Tests were inverted to assert that the engine returns a finite, in-range position for both JDs. This proves the engine handles deep-ancient epochs correctly rather than failing silently.
• Resolved: 2026-05-21

---

Layer 4 - House Singularities

All 53 Layer 4 tests pass as of 2026-05-21.

test_layer4k_body_on_cusp_placement_is_stable_and_deterministic is now executable: house_of() has been implemented in moira.houses, so the former skip has lifted and the test passes on current main.

---

How to Update This Document

When a defect is fixed:

1. Move the entry from Open Defects to a new Resolved section.
2. Add a Fix: line naming the commit SHA and the module(s) changed.
3. Add a Verified: line stating which test now passes and the date.

When a test-design finding is corrected:

1. Update the constant in the test file.
2. Note the correction in the TDF entry and mark it Resolved.

When a new adversarial attack is added to the suite:

1. Run the full suite and update the summary counts at the top.
2. Add any new failures as new DEF or TDF entries.

When a new oracle-hostile audit is added:

1. State the authority surface explicitly (Swiss, ERFA, Horizons, USNO, etc.).
2. State whether the attack is on a direct substrate route or a public datetime path.
3. Declare any route-specific tolerance budget before interpreting a mismatch as an engine defect.
4. Re-run the combined campaign and update both the adversarial and full-campaign counts at the top.