This release upgrades the Huawei Solar communication stack and improves Modbus batching reliability.

Added

• Added get_error_tracker() for tests and diagnostics.
• Added tests covering tracker singleton behavior and reset state.

Changed

• Upgraded huawei-solar from 2.5.0 to 3.0.7.
• Migrated from pyModbus to tModbus.
• Tightened read_registers() error handling so programming errors now fail fast.
• Introduced typed error categories in error_tracker.py.
• Refactored internal error-tracker and register-read code for better maintainability.

Fixed

• Fixed smart batching so requests never exceed the Modbus 125-register limit.
• Fixed storage-unit state-of-charge register mappings.
• Fixed register definition and CallbackAPIVersion import paths.

Dependencies

• Removed: pymodbus, backoff, pyserial-asyncio, pyserial, pytz
• Added: tmodbus, serialx, tenacity

🐛 Fixed

• Modbus connect timeout: setup_modbus() now wraps AsyncHuaweiSolar.create() in asyncio.wait_for() with a 15-second timeout. This prevents indefinite hanging if the TCP handshake stalls. Previously, only the slave-ID auto-detection path had a timeout — the final connection-create call was unbounded (HA-ASYNC-001).

🔧 Changed

• Narrowed exception handling in run_main_cycle: The broad except Exception catch-all was replaced with an explicit RECOVERABLE_EXCEPTIONS tuple (TimeoutError, ConnectionRefusedError, ModbusException). ExceptionResponse from pymodbus 3.x is excluded at runtime, since it is a dataclass and not a BaseException subclass. Truly unexpected errors now propagate up to the top-level handler in main() instead of being silently swallowed. asyncio.CancelledError is explicitly re-raised (CC-CLEAN-002).
• Typed error categories: Introduced a new ErrorType Literal ("timeout", "connection_refused", "modbus_exception") in error_tracker.py. The previous type(e).__name__.lower() stringification (which produced inconsistent names like "timeouterror") was replaced with explicit, canonical error-type mapping (PY-STYLE-003).

🔧 v1.10.3 — Async Safety & Hotpath Fixes

This release eliminates blocking calls in the MQTT publish path and connection setup that could stall the event loop for up to several seconds per cycle.

What changed:

• MQTT hotpath: publish_data() and publish_status() no longer block on synchronous wait_for_publish(). The blocking paho-mqtt wait is now offloaded to loop.run_in_executor().
• Startup: connect_mqtt() replaced its time.sleep(0.1) busy-poll loop with a threading.Event + run_in_executor() pattern. This is the correct primitive since paho callbacks execute in their own thread.
• Shutdown: disconnect_mqtt() is now async; loop_stop() and disconnect() run in an executor instead of blocking SIGTERM handling.
• mypy: Added # type: ignore[attr-defined] for paho-mqtt CallbackAPIVersion.VERSION2 to resolve a type-checker false positive.
• Dependencies: requirements.txt regenerated from uv.lock (pymodbus==3.13.1, pytz==2026.2).

Impact:

• Cycle publish latency reduced from up to ~2s blocking to non-blocking async wait.
• Startup no longer wastes CPU in a busy-poll loop.
• Shutdown is cleaner under SIGTERM/SIGINT.

No breaking changes. All existing configurations and behaviors remain identical.

[v1.10.2] - 2026-06-18

Fixed

• Runtime state encapsulation: Replaced scattered globals (LAST_SUCCESS, BATCH_MODE_AVAILABLE) with a _BridgeState dataclass and a dedicated reset_state() helper — no runtime behavior change, significantly cleaner test isolation
• Poll interval enforcement: Moved cycle wait handling into the main loop path, preserving existing poll_interval timing behavior
• Modbus auto-detection cleanup: client.stop() during failed slave-ID attempts is now bounded by asyncio.wait_for(..., timeout=2.0); cleanup failures are logged instead of silently swallowed
• Filter reset behavior: reset_filter() now resets the singleton state in-place instead of destroying the instance, closing a brief unprotected window during error recovery where a zero-drop could reach Home Assistant
• TRACE logger typing: Replaced dynamic monkey-patching of logging.Logger with a dedicated _TraceLogger subclass

Changed

• Default batch_max_gap remains at 50: Keeps Batch 3 below the inverter's internal 125-register limit after the 1.10.1 adjustment — if you have batch_max_gap: 100 in your config, update it to 50
• Runtime dependency export: requirements.txt is now generated by uv export --no-dev --no-hashes --no-emit-project and includes all transitive runtime dependencies
• Development dependency management: Dev dependencies migrated from optional dependencies to PEP 735 [dependency-groups]
• CI dependency resolution: CI uses uv sync --dev; Docker generation uses --no-emit-project
• CI actions: Updated GitHub Actions and astral-sh/setup-uv to latest versions
• Production push script: Removed dev-only local/update-test scripts from the production push summary; added a requirements.txt guard

Tests

• Added tests/test_version.py to verify the bridge version constant
• Expanded BATS coverage for enable_batching and batch_max_gap environment variable handling
• Refactored test suite around the new _BridgeState helpers and [dependency-groups] setup

v1.10.1 - Fix default batch_max_gap

Fixed

• Default batch_max_gap reduced from 100 to 50 (#XX)
  • The previous default caused Batch 3 to exceed the inverter's internal
    register span limit, triggering a sequential fallback on every cycle
  • With 50, batches stay within the inverter's limits out of the box
  • Users with batch_max_gap: 100 in their config should update to 50

Improved

• Clearer log message when a batch exceeds the inverter limit, with a hint
  to reduce batch_max_gap
• Updated config descriptions (EN/DE) to reflect new default

[1.10.0] - 2026-06-02

Added

• Added smart batch register grouping with new configuration options:
  • enable_batching (default: true)
  • batch_max_gap (default: 100)
• Added intelligent register grouping by Modbus address proximity
• Added graceful fallback to sequential reads when batch reads fail
• Added batch_builder.py with a dedicated BatchBuilder class for batch construction
• Added an extensible foundation for future Modbus batching optimizations
• Added batch timing integration into existing DEBUG performance statistics

Fixed

• Improved handling of optional and unsupported registers during batch reads
• Fixed the known DEV-branch limitation where optional storage-related registers could break batch requests
• Expected to resolve #25

Tests

• Added test_batch_builder.py with 15 tests for batch building logic
• Added test_read_registers_smart_batching_enabled
• Extended test_main.py for the new tuple-based batching API
• Extended configuration loading tests for enable_batching and batch_max_gap

Performance

• Reduced TCP calls from 67 individual reads to typically 3-5 batch requests
• Improved typical cycle time from 17.4s to 3.9-6.7s (~77% reduction)
• Improved optimized cycle time from 15.1s to 4.4s during warm-up, then 3.9s steady-state (~74% reduction)
• Preserved partial performance benefits even when fallback to single reads is required

[1.9.0] - 2026-05-03

Added

• Performance Diagnostics: Detailed per-register timing measurements are now available when log_level: DEBUG is enabled

• Individual register read times with automatic warnings for slow registers (>200ms)

• Statistical summary per cycle: average, min, max, and median read times across all registers

• Top 5 slowest registers are listed in the logs to help identify performance bottlenecks

• Connection establishment time is now logged with detailed host/port/slave information

• Aids in diagnosing network latency issues and SDongle firmware performance variations

• Documentation: Expanded Performance Troubleshooting section in DOCS.md including:

  • Step-by-step guide for diagnosing slow cycle times
  • Network latency testing instructions (ping, traceroute)
  • Reference values for expected vs. problematic performance metrics
  • Firmware update recommendations and known caveats

Tests

• Added test_read_registers_timing_and_statistics to verify timing measurement functionality
• Added test_read_registers_no_slow_register_warning_when_fast to ensure no false positives for slow register warnings

[1.8.5] - 2026-04-28

Fixed

• aarch64 / ARM installation failure (Issue #22): Add-on failed to build
  on Raspberry Pi 4 and other ARM-based systems due to a hardcoded amd64
  base image in the Dockerfile
  • Root cause: ARG BUILD_FROM=ghcr.io/home-assistant/amd64-base:latest
    was used as default, causing Docker to pull an amd64-only image when
    building for aarch64
  • Removed hardcoded default — BUILD_FROM is now fully controlled by
    Home Assistant Supervisor via build.yaml
  • Added missing build.yaml with correct architecture-specific base images
    for all supported architectures (aarch64, amd64, armhf, armv7, i386)

CI

• Fixed test-docker job: added --build-arg BUILD_FROM=... and load: true
  to docker/build-push-action so the built image is available for subsequent
  test steps
• Removed redundant second docker build call in test-docker job