Add MarlGrid matrix observations

[tortoise10101]

Summary

• add semantic matrix and full_matrix observation formats for MarlGrid
• keep existing pixel observations as the default via observation_format="pixels"
• add direct MarlGrid tests for partial and full matrix observation shapes and channels
• replace the .dockerignore symlink with explicit Docker ignore rules
• fix property docstrings that fail the current ruff lint rule

Tests

• make bazel-test BAZEL_TEST_TARGETS=//envpool/marlgrid:marlgrid_test
• uvx ruff check envpool/python/envpool.py envpool/python/protocol.py scripts/coverage_summary.py
• uvx ruff format --check envpool/python/envpool.py envpool/python/protocol.py scripts/coverage_summary.py envpool/marlgrid/marlgrid_test.py
• uvx clang-format --style=file -i envpool/marlgrid/marlgrid.h -n --Werror
• git diff --check origin/main..HEAD

Note: the branch was rebuilt on top of current origin/main after native MarlGrid prestige coloring landed upstream in #419.

[chatgpt-codex-connector]

💡 Codex Review

envpool/envpool/marlgrid/registration.py

Line 50 in f953a37

grid_size=15,

Match the pinned upstream grid size

For this exact public ID, the pinned upstream marlgrid/envs/__init__.py registers MarlGrid-1AgentCluttered15x15-v0 with grid_size=11 despite the name. Setting it to 15 here changes reset distribution, full renders, and full_matrix observation shape relative to the declared oracle; the current alignment test syncs the oracle from EnvPool's rendered frame, so it will not catch this mismatch.

---

envpool/envpool/marlgrid/marlgrid.h

Lines 663 to 664 in f953a37

	if (cell.type != detail::CellType::kEmpty) {
	continue;

Allow agents to spawn on overlapable objects

In the pinned upstream place_obj path, agents can be placed on cells whose object can_overlap() (for example goal and bonus tiles), with the object kept underneath the agent. This check restricts resets to empty cells only, so GoalCycle and cluttered tasks never produce official starts where an agent begins on a goal/bonus tile, changing the reset-state distribution and the initial observations for those seeds.

---

envpool/envpool/marlgrid/marlgrid.h

Line 789 in f953a37

DecayPrestige(&agent);

Decay prestige only after bonus-tile visits

When prestige_coloring=True, this call decays prestige after every action, including turns and moves onto empty cells. The pinned upstream GridAgentInterface.on_step applies prestige_beta only when the moved-onto object is a BonusTile, so after a rewarded bonus visit EnvPool colors fade during unrelated actions while the oracle's prestige color remains stable until the next bonus interaction.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[Trinkle23897]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Another round soon, please!

Reviewed commit: e436ab8e73

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".