Add latest versions of scripts

.cursor/rules/meva-pipeline-quality.mdc

@@ -0,0 +1,14 @@
+---
+description: Quality rules for MEVA QA generation and validation code
+globs: meva/**/*.py,meva/**/*.sh
+alwaysApply: false
+---
+
+# MEVA Pipeline Quality
+
+- Prefer deterministic symbolic QA generation. Naturalization may rephrase questions, but it must not compute or change answers.
+- Use Kitware YAML as the primary annotation source when available; treat NIST/VSet7/MEVID as validation or enrichment sources unless code explicitly says otherwise.
+- Avoid questions that rely on unsupported identity assumptions. Annotation actor IDs are per-camera and can switch across cameras/time.
+- Counting logic must deduplicate multi-camera views of the same event; do not count each camera observation as a separate real-world event.
+- Entity descriptions should prefer the richest verified source available, then fall back conservatively. Avoid leaking vague placeholders like `unknown`.
+- Validation should be category-specific: temporal ordering, spatial distance/proximity, event uniqueness, distractor quality, and answer-option separation have different failure modes.

.cursor/rules/meva-project-context.mdc

@@ -0,0 +1,12 @@
+---
+description: Core MEVA multi-camera QA project context and data boundaries
+alwaysApply: true
+---
+
+# MEVA Project Context
+
+- The MEVA work is a multi-camera video QA pipeline for surveillance slots in the NeuS-QA benchmark. The active project root is `Multi-Camera-Adi/`; home-root scratch files are not source of truth.
+- Treat `/nas/mars/dataset/MEVA/` and `/nas/mars/dataset/MEVA/meva-data-repo/` as external dataset sources. Do not copy large raw videos, annotations, or PDFs into the repo.
+- Keep generated or bulky artifacts out of git unless they are curated examples for showcase purposes. Prefer `meva/examples/` for small representative outputs.
+- Preserve exact answer provenance: QA answers should be computed from symbolic annotations / scene graphs, with LLMs used only for optional text naturalization.
+- Never commit `.env`, API keys, local caches, virtual environments, model weights, raw videos, or NAS-derived bulk exports.

.cursor/rules/project-history-and-postmortem.mdc

@@ -0,0 +1,11 @@
+---
+description: How to use preserved session logs and project history
+alwaysApply: true
+---
+
+# Project History And Postmortem
+
+- Historical session logs live under `docs/project-history/`. Use them as evidence for postmortems, resume bullets, and project storytelling, not as active implementation instructions.
+- When summarizing project impact, distinguish completed production behavior from abandoned V4-V9 experiments and intermediate QA outputs.
+- Prefer concise synthesized lessons over copying old session-log process rules. The old `.github` logging workflow is archived, not active.
+- For recruiter-facing material, emphasize the final V10 MEVA pipeline, deterministic QA design, cross-camera reasoning, validation quality, and engineering tradeoffs.

.gitignore

@@ -181,7 +181,6 @@ cython_debug/
 env
 *.pyc
 output/
-data/
 lm_cache
 .idea
 build
@@ -213,7 +212,7 @@ LLaVA/
 temp/
 InternVL/
 logs/
-data/
+/data/
 llava-video/
 Video-MME/
 VATEX/