Update TTS providers and story markup

README.md

@@ -61,7 +61,7 @@ Environment variables are loaded from `.env`.
 - `OPENROUTER_API_KEY`: API key for LLM command interpretation.
 - `OPENROUTER_MODEL`: OpenRouter model name.
 
-TTS provider settings are configured in the browser options menu and persisted in browser storage. Providers currently include `none`, browser speech synthesis, Kokoro, ElevenLabs, and OpenAI. Production should not assume a universal TTS default; the game or player state selects the active mode, and `none` is the safe fallback.
+TTS provider settings are configured in the browser options menu and persisted in browser storage. Providers currently include `none`, browser speech synthesis, Kokoro, ElevenLabs, OpenAI, and local OpenAI-compatible servers. Production should not assume a universal TTS default; the game or player state selects the active mode, and `none` is the safe fallback.
 
 ## Starting A Game
 
@@ -116,6 +116,17 @@ The train stops at Eibenreith.
 
 Glossary markup is a normal story tag scoped to the paragraph/block it is attached to. The UI finds every matching visible instance of the term in that right-page block and adds a hover/focus note. The tag itself is not displayed, is not sent to TTS, and is ignored by choices and command history. Avoid raw Ink control characters in the explanation; `|`, `{`, and `}` must be escaped in Ink as `\|`, `\{`, and `\}` if they are needed literally.
 
+TTS reading instructions:
+
+```text
+„Ich habe nichts gesehen“, sagt Viktor.
+#tts[Read softly, with controlled unease.]
+```
+
+`#tts[...]` is scoped to the paragraph/block it is attached to and is sent only to providers that support per-request reading instructions. This providerless form is the normal authoring style; `#tts(...)` is equivalent if parentheses read better. Provider-specific forms are also accepted for overrides, for example `#tts[openai](Read softly.)` or `#tts-openai[Read softly.]`. Currently only OpenAI `gpt-4o-mini-tts` consumes the instruction.
+
+Write TTS instructions as concise performance direction: tone, emotion, intonation, pace, accent, or whispering/singing style. Keep the spoken words in the paragraph itself and use the tag only to guide delivery.
+
 Canonical block/media/control tags use Ink-style `#` syntax. In Ink these are real Ink tags. In YAML and Z-code narrative output, leading `#...` lines are parsed by the server into the same structured `StoryTag` objects before reaching the client. The browser only consumes structured `TurnResult` objects.
 
 Tag format:
@@ -127,7 +138,7 @@ Tag format:
 #key:value
 ```
 
-For Ink choices, put choice-local tags under the choice they belong to. Explicit keyboard letters are supported with `# letter[x]`, `#letter[x]`, or the colon form `#key:x`; the client reserves those keys first, then assigns the remaining visible choices from `1` through `0`, then `A` through `Z` in order. `#optional` renders the choice in italic. `# action[name]` or `#action:name` is parsed as a category/template hint for future choice layouts, although the current UI displays all choices in one list.
+For Ink choices, put choice-local tags under the choice they belong to. Explicit keyboard letters are supported with `# letter[x]`, `#letter[x]`, or the colon form `#key:x`; the client reserves those keys first, then assigns the remaining visible choices from `1` through `0`, then `A` through `Z` in visible order. `#optional` renders the choice in italic. `# action[name]` or `#action:name` assigns an invisible action group: group order follows the first appearance of each action tag in the authored list, entries inside each group are randomized, and choices without an action tag are grouped last.
 
 Chapter: