fix: preserve dev fixes in v2 and add grid fuzzing

scripts/regressions/README.md

@@ -0,0 +1,40 @@
+# Release regression exploration
+
+Run the opt-in grid fuzzer from the repository root after installing dependencies:
+
+```sh
+pnpm exec tsx --tsconfig apps/nextjs/tsconfig.json scripts/regressions/grid-fuzz.mts --seed 6545 --steps 1000
+```
+
+Change the seed to explore another repeatable sequence. `--seed` accepts unsigned
+32-bit integers, including zero; `--steps` accepts 1–100,000 transactions. Each
+transaction has one to four move or resize previews followed by commit or cancel.
+The committed state becomes the next transaction's input. The fixture includes
+two root grids and two nested containers with varying widths and row limits.
+
+The runner calls the real v2 grid engine and checks placement identity and
+metadata, integer geometry, bounds, overlap, container ownership, cycles,
+immutable snapshots, repeat-preview stability, rejection retention, and exact
+cancel/commit results. It does not start Homarr, Docker, or a browser. It uses
+the Next.js app's TypeScript config to resolve the engine's existing imports.
+
+A failure exits nonzero and writes the seed, source revision and hashes, initial
+state, and operation transcript to `/tmp/homarr-grid-fuzz`. Use `--output <path>`
+to keep artifacts elsewhere. Re-run the printed command on the same source to
+reproduce it; changing engine or runner code can change a seed's outcome. Reduce
+`--steps` to find the first failing transaction. Artifacts contain synthetic data.
+
+Passing this runner covers v2 engine invariants only. It does **not** establish
+parity with `dev`, browser rendering, save/reload persistence, or database upgrades.
+The next useful differential checks are:
+
+- Clone a seeded `dev` database, upgrade only its copy to v2, and compare widget
+  identities/options/integrations through the category/dynamic-to-container
+  migration (`packages/db/migrations/custom/0004_unify_sections_and_gutters.ts`).
+- Drive the same saved board on both versions around layout breakpoints and
+  compare visible widget counts, errors, containment, and save/reload outcomes.
+  Use separate authenticated browser sessions and version-specific selectors;
+  the existing `scripts/benchmarks/dev-board.mts` markers are v2-specific.
+- Reuse mock integration fixtures from `e2e/lazy-widgets.spec.ts` and geometry
+  checks from `e2e/media-server-layout.spec.ts`, keeping intentional v2 redesigns
+  separate from missing behavior. Pixel equality is not a release criterion.