Status: Accepted · Recorded: 2026-09

Context

ADR 0038 made a 1W identity's burst shape configurable through a tri-state low_power: key, and left unset meaning LEGACY_LONG — the 1024-byte preamble on all four copies — so that no existing installation's on-air bytes moved.

That default cannot work on a VELUX receiver. An awake VELUX receiver does not accept a frame behind the long preamble; it accepts the same frame behind a short one. The effect was isolated on 2W, by a bisection that varied the preamble and the CTRL1_LOW_POWER flag independently, and replicated on 1W: a mains-fed VELUX roller shutter did not move for a burst sent 1024/1024/1024/1024, and moved for both shapes that put any copy at 32 bytes, with CTRL1 held at 0x00 throughout and nothing else changed.

1W carries no acknowledgement, so the hub cannot tell a burst that was obeyed from one that was ignored. The default therefore turns a one-line omission into a silent failure that presents as "pairing does not work", with nothing in the logs pointing at the radio.

ADR 0038 named the evidence that would settle whether the default should change — "a commands-only A/B on an already-enrolled identity" — and issue #74 ran exactly that.

The same reasoning does not extend to Somfy. A mains-fed Somfy awning has been measured taking the long preamble reliably while missing a substantial fraction of short start frames, so a blanket change would risk a regression on the one vendor path this project has validated most.

Options considered

Default everything to ALWAYS_ALIVE (full ADR 0029 parity). Rejected for now: it moves Somfy's on-air bytes with no Somfy evidence, and the measurement above suggests the move could hurt. This remains the intended end state once a Somfy regression run exists.
Default everything to LOW_POWER (one long wake copy, three short repeats). Tempting, since it serves mains and duty-cycled receivers alike and matches a real KLI remote's burst. Rejected: it is not byte-identical to legacy for Somfy either — copy 1 gains CTRL1_LOW_POWER. The 1W MAC spans cmd+data only, so the MAC stays valid, but a receiver filtering on CTRL1 would break, and nothing rules that out.
Resolve the unset default from the manufacturer profile (chosen). The evidence is manufacturer-shaped, so the default is too.
Keep the legacy default and document the workaround. Rejected: the documentation already told VELUX users to set the key, and they still hit the failure, because nothing in the symptom points at a radio setting.

Decision

OneWayWireProfile (oneway_controller.h) gains default_power_class, the shape used by an identity that does not set low_power: at all:

manufacturer:	default_power_class	Why
velux	ALWAYS_ALIVE	Issues #87 and #74: an awake VELUX receiver rejects the long preamble.
somfy, unset	LEGACY_LONG	Byte-identical to before; no Somfy regression run exists.
anything else (profile_is_a_guess)	LEGACY_LONG	A profile nobody has measured never silently acquires a shape nobody tested for it.

This follows the pattern ADR 0031 already established for execute_acei and ADR 0032 for enrollment_classes: a manufacturer profile supplies the default, an explicit YAML key overrides it, and one effective_*() helper is the only place the two are combined.

The identity stores intent, not a shape. OneWayControllerIdentity::power_class_override is a std::optional<OneWayPowerClass>, empty when the key is absent. OneWayPowerClass keeps exactly three enumerators, each a real on-air shape, so oneway_burst_copy_shape() still switches exhaustively with no unreachable case. A fourth "unset" enumerator was avoided precisely because it would make that function accept a value it cannot render.
effective_power_class(identity) is the single resolver, and the only thing the transmitter and the boot log ask. has_power_class_override() reports which side the answer came from.
Resolution stays in C++. Codegen emits the three configured states as three distinct values ({}, ALWAYS_ALIVE, LOW_POWER) and never consults the manufacturer, so the profile lookup lives in one language rather than being duplicated in __init__.py where it would drift from the ACEI and enrollment-class defaults beside it.
The boot log prints the resolved class and its origin — power always-alive (from profile) — since the class alone would not tell a VELUX user whether they had chosen it or inherited it.

Consequences

A VELUX 1W identity that omits low_power: changes shape, from 1024/1024/1024/1024 to 32/32/32/32. For those users this is the difference between a hub that does nothing and one that works; there is no configuration for which the old default was better on that hardware.
Somfy and unrecognised manufacturers are byte-identical to ADR 0038. Nothing about their transmissions changes.
An explicit low_power: still wins everywhere, so a user who pinned a shape keeps it, and a VELUX user with an unusual receiver can still force LEGACY_LONG back.
Unset is now manufacturer-dependent, which is one more thing low_power: means. The key was already tri-state and already differs from the 2W key of the same name; the boot log's (from profile) marker exists so the resolved answer is never a guess.
ADR 0029 parity is still not reached. LEGACY_LONG remains the default for Somfy, still violating the "flag and preamble never disagree" invariant, until a Somfy regression run says otherwise. This ADR narrows that debt to one vendor rather than clearing it.

See also ADR 0038 (the shapes and the tri-state key), ADR 0029 (the 2W precedent), and ADR 0031 (the manufacturer-profile pattern this follows).