§7's spectrum-integrity tolerance is falsified

Contract §7 (Scientific Sanity Checks):

| Spectrum integrity | T2 channel-sum vs T1 counts_total; T5 vs T3 band sums | ≤ 0.1 % (SoLEXS) |

The discrepancy is ~160%, against a stated tolerance of 0.1%. The relationship is not noisy — it is systematic and one-directional.

LOGICALLY IMPLIED: the .lc is not the sum over all 340 PI channels. It is a band-limited light curve. Supporting OBSERVED evidence: the .lc RATE HDU carries NUMBAND = '4' — the SoLEXS pipeline defines four bands — while exposing only TIME and COUNTS, with no band identification.

A search for the contiguous channel range best reproducing the .lc found [40:170] at mean abs diff 38.3 — far better than the full-band 569.6, but still only 16/86,395 exact matches. HYPOTHESIS: the .lc band is defined by an energy interval mapped through a response the archive does not contain (see §2.2: no RMF/ARF exists anywhere in the archive), so it cannot be reconstructed from channel indices alone. This is consistent with, and downstream of, the already-recorded RMF gap.

Root cause of the §7 defect: the same class as CONTRADICTION-001 — I asserted a relationship between two fields from their names and OGIP conventions (HDUCLAS2='TOTAL') without computing it. HDUCLAS2='TOTAL' means "not background-subtracted", not "all channels"; I over-read it.

No implemented behaviour depends on §7. The M-IV parser neither sums channels nor cross-references the .lc. §7 is Milestone VIII's validation protocol, and §3's T1/T2 (Milestone VII) do not require the two to reconcile.

But it changes what M-VII should record: if T1 counts_total (from .lc) and T2 counts (340 channels) are not commensurable, that must be stated in the schema, or a downstream consumer will reasonably assume sum(T2.counts) == T1.counts_total and be wrong by ~2.6×.

The five items below were not adopted. Item 1 (a "tracked diagnostic" ratio) and item 4 (assumption A-10) would have written an unvalidated observation into the contract — the owner correctly refused. Nothing here is in force. Superseded by the r3 deferral note.

1. §7 Spectrum integrity — replace the falsified row with two honest checks:
   • T2 internal consistency: per-second Σ(340 channels) finite ⟺ .lc COUNTS finite (the NaN⟺NaN structural check, which does hold — verified three-way at M-IV).
   • T2 vs T1 relationship: NOT a tolerance check. Record the measured ratio Σ(340)/counts_total per day as a reported diagnostic; flag any day whose median ratio departs from the archive-wide median by >20% as an instrument-configuration change worth investigating. No pass/fail threshold, because the correct value is unknown without an RMF.
2. §3 T1/T2 — add a binding schema note: T1.counts_total (from .lc) is band-limited and is NOT Σ(T2.counts); the observed ratio is ≈2.6 on the reference day. Consumers MUST NOT treat them as interchangeable.
3. §2.1 — record that NUMBAND='4' and the band definition are not recoverable from the archive (no RMF), promoting A-6 from "captured, not interpreted" to an explicit acquisition dependency.
4. §8 — new assumption A-10: the .lc-vs-spectrum ratio (median 2.60) is VERIFIED on the reference day only. Milestone VIII must report the distribution across all 436 archives; a bimodal or drifting ratio would indicate a band or gain change and is a scientific finding.
5. 0.5.3 acquisition — the RMF/ARF request already ranked #2 should additionally request the .lc band definition; without it the .lc cannot be related to the spectra at all.

Approve. As with r1/r2 this strengthens: it replaces a fabricated tolerance with a measured diagnostic, and it closes a real misuse path (sum(T2) == T1) before any consumer depends on it. It also converts an unexplained 2.6× discrepancy into a tracked, archive-wide observable.

Deliberately NOT proposed: widening the 0.1% tolerance to fit the data. The relationship is not approximately-true-with-slop; it is a different quantity. A loosened tolerance would be exactly the "make it work" failure the implementation philosophy forbids.

1. Do the two products measure different physical quantities?
2. Does onboard processing explain the difference?
3. Does official ISSDC documentation resolve the discrepancy?
4. Does any reproducible mapping exist between them?

Until answered, no v2 code, schema, or artifact may assume, assert, or rely on any relationship between T1.counts_total and T2.counts.

Milestone IV COMPLETE: 93/93 tests, 0 xfail, 0 skip, zero deviations. Real D1 .pi: (86400, 340), CHANTYPE='PI', DETCHANS=340, exposure 1.0 s/spectrum, channel map constant 0..339 (validated then collapsed, F-08), 5 all-NaN spectra at [0, 5, 30072, 30078, 83951] — a three-way agreement with the .lc NaN set and the GTI-excluded set. V-PI-3 (.pi TSTART[0] == .lc TSTART) passes. Parse: 0.5 s, 2.10 GB peak RSS — confirming §2.2's streaming mandate is a real constraint, not a stylistic one. Milestone V not started.