feat(discovery): inverter records carry entities, capabilities and Predbat-named ratings (observe-only, piece 1)

[springfall2008]

Piece 1 of moving automatic inverter configuration into the discovery coordinator. Observe-only: every inverter reporter's discovery record now carries enough to rebuild its INVERTER_DEF row and to reproduce what its automatic_config() binds, but nothing configures from it yet. No automatic_config(), inverter.py or write path changes; each reporter's automatic_config() was checked byte-identical before commit.

Design: docs/superpowers/specs/2026-09-24-discovery-inverter-record-vocabulary-design.md (decisions D1-D15). Plan: docs/superpowers/plans/2026-09-24-discovery-inverter-record-vocabulary.md.

What changes

• Record vocabulary (coordinator.py, schema_version 1 -> 2)
  • entities: every setting automatic configuration binds for the device, keyed by Predbat setting name. Each descriptor needs access (rw Predbat writes / r Predbat reads) and exactly one of entity_id or a fixed value; invert replaces the *_power_invert settings.
  • capabilities: a dict of the seven INVERTER_DEF behaviour keys (support_charge_freeze, support_discharge_freeze, support_feedin_first, can_span_midnight, charge_discharge_with_rate, charge_control_immediate, target_soc_used_for_discharge) -> bool. Replaces the old token list.
  • ratings under Predbat setting names: inverter_limit, export_limit (configured grid export cap), import_limit, battery_rate_max, soc_max, battery_min_soc. Only figures the device reports - a value Predbat derives or the user overrides is an entity, not a rating (D14).
  • inverter_definition(record, write_and_poll_sleep, base=None): a pure function that rebuilds an INVERTER_DEF-shaped dict from a record. The has_* presence flags, time formats, SoC units, charge-control mode and load-entity count are derived from the entity map rather than stated twice. Never mutates INVERTER_DEF.
• Seven reporters converted: Sunsynk, Deye, AlphaESS, Fox, Solis, GE Cloud, GivTCP. Records describe their own device (D10), ignore user opt-outs like automatic_ignore_pv (D11), and PV-only devices carry read-only pv_power/pv_today (D12).
• tests/discovery_contract.py: shared checks every reporter's tests run - completeness (the record alone, with no row as a base, rebuilds its INVERTER_DEF row) and agreement in both directions (the record binds exactly what the unchanged automatic_config() binds, apart from named per-device/opt-out extras).
• One INVERTER_DEF change: GEC and GEE soc_units "kWh" -> "%" (GE Cloud binds soc_percent; inv_soc_units is never read). GE-family clock_time_format is left alone because those rows also serve the GivTCP add-on and GE Cloud templates; the GivTCP/GE Cloud records describe the ISO format their sensors really publish, with a named exception in the completeness test (D13).
• ComponentBase.WRITE_AND_POLL_SLEEP (2; GivTCP and GE Cloud 10) - read only by tests until piece 3.
• Docs: docs/discovery-catalogue.md and the catalogue design spec updated for schema 2; a debug-journal entry on a worktree trap (running the harness in a fresh worktree without manifest.yaml self-updates over tracked files).

Visible effect: the discovery section of debug dumps and the /discovery page show the new shapes. GivTCP's entity map now lists only settings automatic configuration binds (so fewer entities than before).

Testing

• ./run_all --quick: all tests pass (4 slow skipped, ~73s).
• Final whole-branch review by a fresh reviewer: ready to merge, no Critical/Important findings; --quick --shuffle over six seeds, all green.
• Per reporter: completeness, agreement both ways, two-device and mixed-fleet tests; Fox/Solis also pin that not-yet-read devices carry nothing.

Known follow-ups (not blocking; items 1-2 should land before piece 3)

1. Deye/AlphaESS drop an energy-counter binding on a poll that omits it; Sunsynk keeps bindings once seen - pick one policy.
2. inverter_definition() prefers soc_kw over soc_percent; inverter.py prefers soc_percent (no reporter binds both today).
3. Fox export_limit/import_limit ratings accept only numeric setting values; if live settings/get returns strings the rating is silently absent - needs a live capture.
4. Test-strength nits: redaction test uses value: "0" not "True" and doesn't check invert; capture_automatic_config() doesn't restore a pre-existing instance set_arg; AlphaESS/GE Cloud completeness loops don't assert a record was checked.

Next: piece 2 (SolaX, Sigenergy, Enphase, Teslemetry, Gateway reporters), then piece 3 (the coordinator configures from records; the spec lists what piece 3 must handle).

🤖 Generated with Claude Code

[mitulgovind]

Question about how the later pieces will handle inverters that have no discovery reporter.

I'm running a LuxPower LXP 3600ACS through the LuxPower HA integration, configured manually in apps.yaml (Predbat v9.3.1, HA add-on). Discovery picks up my Octopus meters and both Solcast sites with no conflicts. inverters: is empty because LuxPower has no discovery reporter, and resulting_config still shows num_inverters: 1 from apps.yaml:

inverters: []
observations:
  conflicts: []
  resulting_config:
    num_inverters: 1
    inverter_type: '#5f01158e'

Everything works fine today since this is observe-only. My question is for when discovery starts driving configuration: will a manual apps.yaml inverter with no discovery reporter always take precedence, and could an empty inverters: [] ever be read as "no inverter found" and override it?

Happy to share the full (anonymised) discovery output if it helps.