docs: move research dossier into docs/ and update cross-references

docs/research_log.md

@@ -0,0 +1,25 @@
+# Research log
+
+## 2026-10-31 — independent verification and first prototype
+
+**Literature milestone.** Searched current publisher, proceedings, preprint and author-repository sources independently of previous discussion. Recorded 34 unique bibliography entries (including one pending primary-access NP reference), with overlapping project placements. Main corrections: EDGRU has a 2026-10-27 early-article record; HyperTea has a 2025 preprint and 2026 journal record; MIST is a 2026 TIP project, not merely an unspecified dataset; SD Loss is AAAI 2025 though its preprint is December 2024. DeepPro preprint and journal author order/status differ. Some publisher pages were blocked, so primary indexed excerpts or explicitly marked secondary metadata were retained. No citation counts were inferred.
+
+**Failed hypotheses.** Proximity-conflict constraints can force deletion of true adjacent targets. Independent-set optimization cannot reject isolated positive-score clutter by itself. Pixel thresholds need not make object FP monotone. Generic size-sensitive loss and temporal graph fusion have substantial prior art. A training constraint is not a test-time guarantee.
+
+**Decision.** Revised B is the first falsification target; C second; original A third. No novel theorem or publication-ready performance claim. Full-text novelty gates remain open.
+
+**Analytical milestone.** Derived matching-deficiency monotonicity for frozen candidates; canonical assignment gives a conservative smooth FP upper bound; iid finite-grid tail-sum CP intervals give a population expected-count certificate. Derived explicit sample-size limitations. All use known combinatorial/statistical tools. Proved assumptions and counterexamples are in `docs/mathematical_analysis.md`.
+
+**Implementation milestone.** Created original compact U-Net/loss/gate training code, fixed-pool metrics/calibration, data audit, scripts and ablation matrix. Pure CPU mathematical and integration checks pass. 24 tests discovered: 22 pass; 2 PyTorch gradient tests skipped. Syntax compilation passes. No dataset/GPU experiment ran; PyTorch and an exposed NVIDIA device are unavailable. The fixture pipeline exercises reject-all handling and certificate invalidation, not real IRSTD accuracy.
+
+**Actual numeric checks.** With M=101, δ=.05 and α=.01, zero observed events require n=758 for the any-alarm CP bound. Under K=64 tail-sum count calibration the conservative zero-event requirement is n=75,321. At n=80 the any-alarm upper bound is about .09075, despite zero observed events. These are analytical/sample-size checks, not detector results.
+
+**Unresolved.** Some DOI/author/code fields and dataset license scopes remain unverified. No archives were downloaded; actual size distributions, duplicate audit and annotation quality remain unknown. MIST GPU adaptation is untested. Gate-training gradient/runtime checks await PyTorch. Obtain full texts of the closest loss/NP methods before treating revised B as novel; obtain adequate independent exposure before claiming stringent false-alarm guarantees.
+
+**Next concrete milestone.** Complete full-text/code audit and data rights/manifest checks, then measure whether any useful certificate is feasible before spending substantial GPU time. The prepared scripts do not imply a continuing background job.
+
+## Follow-up audit and calibration refinement
+
+Corrected the cutoff to the supplied environment date, October 31 UTC; future November coverage remains impossible. Verified EDGRU code listing and commit, inspected SLS/SD loss source, and recorded implementation caveats without copying upstream code. Read LTT's bounded-test and fixed-sequence sections; added the known fixed-sequence baseline and its proof assumptions.
+
+Executed 28 tests: 26 passed, 2 PyTorch tests skipped. An initial rerun failed two subprocess imports because PYTHONPATH was omitted; rerunning with the documented PYTHONPATH=. command passed. This was a test invocation issue, not a detector experiment. Four new tests check exact zero-event sample boundaries, stopping without restart, count calibration and support endpoints. Fixed-sequence zero-event counts are 299 (any risk) and 19,172 (count risk, K=64). No real dataset, training or GPU profiling ran.