ADR-0005: Universal IR — code database & rendering

IR encoding is pure and in tv-core::ir (NEC family, 38 kHz, header + 32 LSB-first bits + stop), so it is fully unit-tested with no hardware. Ships a samsung_tv code set (Samsung’s 0xE0E0 custom code); other brands slot in as new tables.

Transmit is an Android-only Tauri plugin (IrPlugin.kt → ConsumerIrManager.transmit). The Rust ir_send_key command computes the pattern in tv-core and forwards it; on desktop (no emitter) it returns a clear error.

android.permission.TRANSMIT_IR is required — found on-device: without it ConsumerIrManager.transmit throws SecurityException ("permission never granted" warning). Added to the manifest + a consumerir uses-feature.

IR devices are client-only pseudo-devices in the UI (id ir:<code_set>, vendor ir); they skip pairing/persistence and route key presses to ir_send_key instead of a TvBackend.

The code database embeds a curated slice of Flipper-IRDB (crates/tv-core/irdb/<Brand>/.ir, 19 TV brands, 87 code sets) and renders each .ir entry’s (protocol, address, command) to timings. The UI is a *find-code wizard (send POWER → "did it react?" → next/save), because Android’s ConsumerIrManager is transmit-only (no way to learn a code from the real remote).

Rendered protocols: NEC, NECext, Samsung32, SIRC/15/20, RC5/RC5X (+ raw passthrough).

RC5/RC5X are now rendered (Manchester bi-phase, 36 kHz). 14 bits MSB-first (S1=1; S2=1 for RC5 or !cmd bit6 for RC5X; toggle T; 5 addr + 6 cmd bits), half-bit 889µs. We build half-bit levels then coalesce equal-level runs into durations; the pattern is forced to start (and end) on a mark by folding the leading/trailing space into the inter-frame gap. This unblocks Grundig/Philips volume/power, which previously rendered nothing.

RC5 toggle: a module-level AtomicU32 flips on every rendered frame, so two taps of the same key send different-toggle frames → the TV scores each tap as a fresh press = exactly one step (no runaway auto-repeat). Deterministically unit-tested; the actual RC5 waveform on a real TV is unverifiable without an IR blaster.

Over-captured raw frames are trimmed to the first frame on parse: some .ir raw entries were captured with the button held, so data: holds a frame + inter-frame gap + repeat, making the TV register multiple presses (Grundig volume stepped 2× per tap). We cut at the first OFF slot (odd index) whose duration exceeds 15 ms — a real inter-frame gap; intra-frame spaces are < ~2 ms — leaving one clean frame ending on a mark (unchanged if no such gap, never trimmed below 3 slots).

Model-scan matching is bidirectional (find_by_model): the query is a whole OCR blob off the back-of-set sticker, so we match if the normalized id/label contains the query (short typed model) OR any of a set’s model tokens (filename minus brand prefix, split, normalized, len ≥ 4) is contained in the query (a blob mentioning the model). Fixes OCR finding nothing because the old code only matched id/label containing the query.