feat: 导入 IRSTD 研究代码与文档（小目标检测）

- 新增 irstd 核心包、scripts、tests、configs
- 新增研究文档：literature_review / mathematical_analysis / novelty_matrix / experiment_plan / research_decision / research_log / manuscript_outline
- 新增 evidence、results 数据与审计产物
- 新增 pyproject.toml、requirements、.gitignore

novelty_matrix.md

@@ -0,0 +1,38 @@
+# Novelty matrix
+
+Reference IDs resolve to `literature_review.md` and `evidence/papers.json`. “No certificate established” means none was established in this audit; it does not assert absence from an unread full text.
+
+## Closest three per original proposal
+
+| Project | Closest prior method | Already covered | Potential remaining contribution | Reviewer objection / required test |
+|---|---|---|---|---|
+| A | Candidate graph classification, Deng et al. 2026 (A2) | Propose regions, graph features, reject false candidates | Correctly justified exclusivity with measured close-pair protection | Merely graph postprocessing; compare to independent MLP gate and NMS |
+| A | EDGRU (A1) | Edge-guided graph reasoning and clutter discrimination | Explicit candidate feasibility rather than feature affinity | Hard edges may delete true targets; require counterexample-driven stress tests |
+| A | CS-ViG-UNet (A3) | Sparse graph features in an efficient detector | Sparse candidate selection with provable assumptions | GNN may add cost without gain; parameter-matched nongraph baseline |
+| B | SD Loss (B2) | Size-dependent localization/scale loss | Count-budget-aware object selection and measured calibration utility | Another size weight is incremental; remove all size weights as an ablation |
+| B | AC-SLSIoU (B4) | Hard-negative margins and false-alarm weighting | Match-defined FP bound rather than negative-pixel penalty | Why not tune the existing loss and threshold? Give all baselines identical calibration |
+| B | CG-IRNet (B5) | Instance structure and confidence-based background suppression | Explicit held-out expected-count certificate with sample-size limits | Certificate might select empty output; report abstention rate and recall |
+| C | HyperTea (C3) | Temporal graph/high-order enhancement plus alignment | Candidate-path decisions with explicit uncertainty and computation budget | Hypergraph→graph may be simplification only; same input window and budget |
+| C | MISTNet (C1) | Irregular/fast motion, implicit compensation | Explicit sparse trajectory inference | Motion assumptions fail on MIST-Hard; oracle-edge recall and gate-miss audit |
+| C | DeepPro (C2) | Long-term weak-evidence accumulation | Sparse path-conditioned evidence under camera motion | Long profiles may be cheaper and better; equal temporal context comparison |
+
+For revised B, **LTT (B13), CRC (B14), and differentiable NP learning (X1, primary access pending)** are conceptually closer than many IRSTD architectures. They must appear prominently in a manuscript. B10 already motivates stronger IRSTD evaluation. For C, classical DP–MHT–TBD (C8) prevents a claim that trajectory optimization is new. A9 prevents a naive close-target exclusion story.
+
+## What can and cannot be claimed now
+
+| Proposed claim | Assessment |
+|---|---|
+| First graph-based IRSTD / first graph false-alarm suppression | Contradicted by verified prior art |
+| First size-sensitive or false-alarm-aware IRSTD loss | Contradicted by verified prior art |
+| First temporal graph detector / first weak-evidence temporal accumulation | Contradicted by verified prior art |
+| First differentiable constrained detector | Not supportable; NP and risk-control prior art |
+| New general statistical calibration theorem | Not claimed; CP, union bounds and LTT are established |
+| Matching-consistent candidate training plus calibrated object-count selection for IRSTD | Plausible application-level research hypothesis; novelty not certified |
+| Better recall at the same statistically certified count risk | Strong experimentally testable claim; currently no evidence |
+| Guaranteed retention of arbitrary close targets | False absent candidate/source identifiability and safe edges |
+
+The useful mathematical mechanism is that the frozen-pool deficiency |S|−ν(S,G) permits coherent object-level calibration and a simple differentiable upper bound. Applying known calibration alone would likely be an incremental paper. Publishability requires a nontrivial, repeatable gain from training under that mechanism, clear evidence on real clutter, and comparison with calibrated existing models.
+
+## Audit update
+
+EDGRU code is now accessible (see evidence/upstream_audit.md), strengthening its role as an executable comparison candidate pending license and environment checks. SLS/SD source inspection confirms that their scale/localization reductions must be reproduced exactly. Fixed-sequence risk testing is explicitly established in LTT; replacing the conservative calibration bound is a baseline improvement, not an additional novelty claim.