Revise scientific conclusions: global optimality, rate frontiers, and source errors

[GoGoKo699]

Scientific changes (version 1.2.0)

This is a substantive revision, not another software-only update.

• Proves that the original special-outcome all-H decoder is optimal over all one-cell quantum POVMs under the specified uniform-secret priors.
• Proves that independently refreshing full secrets within a fixed parity class makes the same product decoder globally optimal for any number of samples; that baseline repair removes the collective accuracy advantage.
• Gives exact achievable GLOBAL/SEP/PPT/LOCC confidence-versus-total-answer-rate frontiers for the ideal source and a precisely specified, externally calibrated phase-noise family. The ideal gap disappears at total answer rate q >= 1/N; increasing N lowers the useful rate, not the maximum confidence gap.
• Supplies exact decomposable PPT operator certificates: an ideal violation requires an NPT effective conclusive effect. The sequential operational null is not broadened to arbitrary PPT instruments.
• Distinguishes common secret-independent separable source channels (original null preserved) from small arbitrary preparation leakage (75% ceiling can fail under abstention).
• Proves a conservative source-error allowance and fixed-total-round statistical certificate, with explicit conditional calibration, side-channel, separable-source/memory, and stopping assumptions.
• Makes the ideal sequential instrument definition and full-history filtration explicit and adds a finite-horizon supermartingale proof.

Provenance and limits

New theorems are dated 7 September 2026. The August correction Markdown/TeX/PDF, five earlier result records, six historical scripts, original release, and August GPT-5.6 Pro credit are preserved. The original shared-secret p_B verification implication remains withdrawn. No universal-computation, speedup, resource-minimality, hardware calibration, or independent-peer-review claim is made. General fixed-inconclusive-rate discrimination and the related Lee–Bae framework are credited; a noisy-input reference locator is corrected.

Validation

• 98 local tests passed, including 1268 individual result-field corruption checks.
• 96 independent GLOBAL/PPT linear programs agree with the exact frontier, starting from raw state preparation rather than the analytical constructor; maximum observed absolute objective error 1.39e-16.
• All 6 scientific result records, 71 file checksums, and 6 historical Git hashes passed.
• The uploaded Git tree exactly matches the tested local tree: bceb165d695686593425c2a87605705e71aa2311.
• CI runs Python 3.10–3.13 plus the independent optimization audit. SciPy is optional audit-only, not a runtime dependency.

The new primary scientific note is correction/SCIENTIFIC_REVISION_2026_09.md. The landing page, status map, claim ledger, witness documentation, code, results, validation, and citation metadata are updated together.