Update AssemblyAI skill: Sync keyterms_prompt/timestamps, async language defaults, Voice Agent voice roster + Sessions API

[dlange-aai]

Scheduled docs-sync run against the current assemblyai-docs api-reference/ + root OpenAPI/AsyncAPI specs (docs repo at 2026-07-20). Covers changes merged since the last skill update (2026-07-10). Stacked on update-assemblyai-skill-2026-07-10 (PR #18).

Correctness fixes

• Sync STT keyterms field is keyterms_prompt (not word_boost). The docs now document keyterms_prompt as the field, with keyterms / word_boost accepted as aliases — "provide only one of the three." The skill previously asserted the opposite (Sync uses word_boost, not keyterms_prompt). Fixed in SKILL.md (Sync config fields + Common Mistakes) and api-reference.md (config table + example curl). (docs #216)
• Sync STT word timestamps are now opt-in via config.timestamps (default false) — start/end are omitted from each words object unless set. The skill showed them as always present. Fixed the response shape/example in SKILL.md and api-reference.md. (docs #210)
• Async language_code / language_detection defaults changed. Neither defaults to en_us / false anymore: omitting language_code auto-runs detection, the two are mutually exclusive, and language_confidence_threshold requires detection on. Updated the param table, the Language Detection section, and added a gotcha in SKILL.md. (docs #231)
• remove_audio_tags now defaults to "all" (all inline annotations stripped by default; pass null to keep them). Enum all/speaker/null documented. (docs #238)
• Voice Agent stored-agent REST voice is an object {"voice_id": "..."}, not a bare string (the WebSocket session.output.voice remains a plain string). Corrected api-reference.md §17 + added a Common-Mistakes row.

Voice roster refresh (July 2026)

• The old roster (ivy, james, tyler, winter, bella, david, arjun, …) is now deprecated (still functional). New current roster: English anna/alba/charles/eve/george/jane/jean/mary/michael (🇺🇸), paul/vera (🇬🇧); language-specific giovanni (IT), lola (ES), juergen (DE), rafael (PT), estelle (FR). Officially supported output languages are now 6 (EN/IT/ES/DE/PT/FR). Updated voice-agents.md and moved examples off the deprecated default ivy. (docs #214, #229)

New endpoints / features documented

• Sync GET /warm connection pre-warming endpoint (unauthenticated, idempotent; warm right before transcribing). (docs #217, #221)
• Voice Agent Sessions API — GET /v1/sessions, GET /v1/sessions/{id}, DELETE /v1/sessions/{id} with recording/timeline/metadata artifacts (pre-signed URLs). (docs #226)
• Webhook auth header constraints (length/charset, mutual + webhook_url requirement) added. (docs #238)
• Param dependency rules documented: redact_pii→format_text, speaker_labels→punctuate, sentiment_analysis→punctuate, speaker_options/speakers_expected mutual exclusivity, redact_static_entities limits. (docs #235, #236, #237)

Verified already-current (no change)

• LLM Gateway model list matches available-models.mdx exactly (Claude Opus 4.7/4.6/Sonnet 4.6/…, GPT-5.5/5.2/…, Gemini 3.5 Flash/…, Qwen3, Kimi K2.5).
• Streaming voice_focus/voice_focus_threshold (default 0.7) and previous_context_n_turns default 5 already accurate.
• redact_pii_sub value entity_name already correct.

Files changed

• SKILL.md
• references/api-reference.md
• references/audio-intelligence.md
• references/voice-agents.md