fix(inverter): restore calculated capacity sensor after restart

[Speshman]

Fixes #5222.

After a restart, battery_size_tracking() recovers today's measurement and trimmed mean from recorder history, but skips its only sensor-publishing branch. Re-publish soc_max_calculated from the recorded measurement only if the live entity is missing. A present entity is left alone, and the daily battery-size calculation is not rerun.

Extend the existing restart regression test to assert that the sensor returns with its state and history and is not re-published on subsequent cycles. Add a case where today's recorded measurement is None, preserving that failed day while restoring the sensor from earlier valid measurements.

Validation: ./run_all --test find_battery_size passed with an isolated empty secrets file; targeted whitespace, encoding, Ruff and Black hooks passed. The full quick run could not be validated in this environment because its secrets-loading test assumes no external Home Assistant secrets file, while this container exposes one. The all-hooks run also hit npm EALLOWGIT initializing markdownlint; those unavailable Node hooks were skipped in the targeted Python hook run.

[springfall2008]

Automated review failed to complete for this PR - see the triage bot's logs for details. The run exited cleanly but posted nothing, so the review step itself did not complete. Not retrying automatically; remove BOT_FAILED and re-add BOT_REVIEW to try again.

[Speshman]

Thanks for checking. Since the code-quality check passed and the review bot exited without posting, could you share the relevant redacted log lines or what prevented the review from being posted? That would help us decide whether a retry is worthwhile.

[springfall2008]

Automated review from the triage bot (/code-review high), posted manually by a maintainer because the bot's own posting step did not run.

Reviewed at head 7217c0df. The fix works for the #5222 scenario. After an HA restart, the periodic update_states() wipes state_data, the elif republishes via set_state (which also refreshes state_data), and the next cycle skips, matching the tests' second-cycle guard. No neighbouring tests in the shared fixture are disturbed.

Findings 1 and 2 are narrow real defects. Findings 3 and 4 are test-coverage and efficiency cleanups. Nothing was found in the removed-behaviour, cross-file or conventions passes.

[Speshman]

Updated in a7d1246 for review points 1–3: the recovery path keeps the already-recorded planning mean (including all-None history), validates recorded history before calling the publisher so malformed entries cannot abort the inverter refresh, and tests that the None-today case does not republish on the next cycle. New regressions cover the all-None one-cycle scaling case and malformed history. With invalid history, the sensor is left missing rather than publishing corrupted data; planning still uses the recorded state.

On point 4: the duplicate reads happen only on the missing-entity recovery cycle. Eliminating both cleanly would require passing cached state/history through or restructuring the publishing helper, so I kept this correctness fix narrow. Happy to consider a separate efficiency follow-up if useful.

The isolated find_battery_size suite, applicable Python hooks and CSpell pass locally. The full quick suite is not validated in this environment because its secrets-loading test assumes no external Home Assistant secrets file.