feat(mesh): session singleton + [mesh] optional dependency

[cagataycali]

Summary

Add strands_robots/mesh_session.py — a thread-safe, ref-counted Zenoh session singleton that will be shared by all Mesh instances in the same process.

Part 1 of 6 for v0.4.0 Zenoh mesh networking. See implementation plan.

What's in this PR

New files

File	LOC	Purpose
strands_robots/mesh_session.py	211	Session lifecycle, MeshConfig, auto-mesh
tests/test_mesh_session.py	258	22 tests, 98% coverage, all mock-based

Modified files

File	Change
pyproject.toml	Add [mesh] extra (eclipse-zenoh>=0.11.0,<0.12.0), add to [all], mypy override

Design decisions

1. Module, not subpackage — mesh_session.py lives at strands_robots/ level. The mesh/ subpackage comes in PR 2 when there are multiple files to organize.

2. MeshConfig(frozen=True) — Immutable after creation for thread safety. Reads ZENOH_CONNECT, ZENOH_LISTEN, STRANDS_MESH_PORT from env.

3. Auto-mesh — When no env vars are set, the first process on a host listens on tcp/127.0.0.1:7447; subsequent processes connect as clients. Zero config required.

4. STRANDS_MESH=false kill switch — Disables mesh globally before any zenoh import.

5. Lazy import — eclipse-zenoh is imported via importlib.import_module() only when get_session() is called. No import-time dependency.

6. Version pin >=0.11.0,<0.12.0 — eclipse-zenoh 1.0 has API breaks in subscriber patterns. Conservative pin for safety; broadened after testing.

7. Return Any — get_session() returns Any (not zenoh.Session) to avoid forcing zenoh as an import-time dependency. Consistent with how the codebase handles other optional deps (torch, lerobot).

API

from strands_robots.mesh_session import get_session, release_session, MeshConfig, session_info

# Acquire (ref +1)
session = get_session()

# Release (ref -1, closes at 0)
release_session()

# Explicit config
session = get_session(config=MeshConfig(connect=("tcp/10.0.0.5:7447",)))

# Diagnostics
info = session_info()  # {'active': True, 'refs': 1}

Test results

22 passed in 0.43s — 98% coverage on mesh_session.py
250 passed (full suite), 6 skipped, 0 failures

What's next

PR	Scope	Depends on
PR 1 (this)	Session singleton + [mesh] extra	—
PR 2	Mesh class + presence + peer discovery	PR 1
PR 3	tell/send/broadcast + response correlation	PR 2
PR 4	publish_step + subscribe + on_stream	PR 2
PR 5	emergency_stop + safety audit log	PR 3
PR 6	Wire mesh into Robot.__init__	PR 5 + v0.3.9

Prototype provenance: dashboard-zenoh-fixes:strands_robots/zenoh_mesh.py (691 LOC, battle-tested with 2 SO-100 arms, 10 simulated G1 humanoids, Mac↔Jetson Thor Wi-Fi at 50Hz).

---

🤖 AI agent response. Strands Agents. Feedback welcome!

[cagataycali]

Closing as duplicate of #101 — three concurrent mesh-session PRs were opened simultaneously (#100, #101, #103). Consolidating on #101 which has the most complete scope (4 files, 801 LOC incl. PeerInfo registry). If reviewers prefer pieces from this PR, we will cherry-pick before merging #101.

Canonical PR: #101
Tracking issue: #98