fix(cloud): azure/edge silent zero-audio — bare <speak> envelopes, <mark>, and zero-audio-as-success

[willwade]

Found while debugging AACTools/VoiceGarden-SPD#1 (speech-dispatcher SSML produced no audio, reported as success). Two Azure/Edge failure modes complete the turn normally (turn.end / HTTP 200) but synthesise zero audio, with no error anywhere:

1. Bare <speak> envelope → zero audio

Azure/Edge accept an envelope missing version/xmlns/xml:lang — the turn completes cleanly and nothing is synthesised. xml:lang is the required missing piece. That is exactly what speech-dispatcher's index-marking wrapper sends, so every SSIP client using SSML mode hits this.

Fix: normalize_ssml_envelope() completes the envelope before the request goes out (WS + REST paths), deriving xml:lang from the voice name (en-GB-SoniaNeural → en-GB), same heuristic as build_azure_ssml. Present attributes, custom namespaces, tag-name case and the rest of the document pass through verbatim; plain text and envelope-less fragments are untouched.

2. SSML <mark> element → zero audio

Azure/Edge don't support <mark>; one anywhere in the document silently zeroes the whole utterance. speech-dispatcher injects <mark name="__spd_N"/> around every pause, so this is the same repro from the other direction.

Fix: strip_unsupported_marks() drops the (empty, unspoken) <mark>/</mark> tags on the azure/edge SSML path. Consumers needing positions should use word-boundary events, which already exist.

3. Zero audio reported as Ok(()) (all cloud engines)

The wrapper was the only layer that knew a turn produced nothing, yet returned success — every consumer had to detect "no bytes" independently. Now:

• WS path (edge + azure-with-boundaries): wire audio frames counted; clean turn.end with zero bytes → Err("… synthesis completed with no audio").
• All REST paths (streaming MP3/PCM, Google audioContent, ElevenLabs audio_base64, raw body): decoded/delivered bytes counted; zero → Err("… synthesis returned no audio").

This is what surfaced failure #2 during verification — the error path works.

Verification

• New unit tests: envelope completion (order, gaps-only, case, self-closing, fragments untouched, composition with inject_voice_if_missing), voice_lang, mark stripping (speechd shape, paired tags, <market> untouched, unterminated left alone).
• Live against the Edge endpoint via cargo run --example edge-bare-envelope: plain text, bare <speak>, and <speak>+<mark> all synthesise audio with word boundaries (previously: 0 bytes, Ok(())).
• cargo fmt, cargo clippy -D warnings (cloud and system,cloud feature sets), full test suite green.

Consumers: VoiceGarden-SPD currently works around #1/#2 module-side (AACTools/VoiceGarden-SPD#3); once this merges, the module's ensure_ssml_document becomes a harmless no-op and can be simplified.