v0.18.0: cache-key migration, audio input, cached_only, service-tier pricing (#45)

* Fix duplicate additionalProperties in JSON schema serialization

The schema had duplicate `additionalProperties` keys: one from
SchemaFormat's struct field and another from OpenAiTransform. This
produced malformed JSON that OpenAI started rejecting.

Fix: OpenAiTransform now only adds `additionalProperties` to
object-type schemas (not primitives like string/integer), and the
redundant field is removed from SchemaFormat. Also adds a unit test
to prevent regression, plus an OpenAI integration test.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Fix duplicate additionalProperties in schema + add cache migration

The schema had duplicate additionalProperties keys: one from
SchemaFormat's struct field and another from OpenAiTransform. This
produced malformed JSON.

Fix:
- OpenAiTransform now only adds additionalProperties to object-type
  schemas (not primitives like string/integer)
- Removed redundant additionalProperties field from SchemaFormat
- Added legacy cache key migration so existing caches are found under
  the old key format and copied to the new key

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Remove unnecessary Gemini-specific additionalProperties stripping

The schema fix (only adding additionalProperties to object-type
schemas, no duplicates) makes the Gemini-specific stripping in
chat_uncached redundant.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Add audio input support and Gemini transcription test

- Add InputAudio variant to ChatMessageContent for sending audio
  (wav/mp3) as base64-encoded data in chat completions
- Add InputAudio struct with wav() and mp3() convenience constructors
- Add base64 dependency
- Move harvard.wav to test_fixtures/
- Add gemini_audio_transcription integration test

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* gemini audio

* add cached_only option

* Bump to 0.18.0: service-tier pricing, new models, cost signature change

Adds flex and priority tier price tables, new gpt-5.4-mini/nano entries,
and threads service_tier through cost() so ChatClient::cost picks the
right table. Pricing snapshot refreshed to 2026-03-17.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Clippy: collapse legacy-cache-key lookup to ? operator

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Update test expectations for cleaned-up schema format

- multiple_optional_fields: model now returns "Elon Reeve Musk" (full legal name) under the new schema
- optional_fields_varying_presence: the model was filling `title` with the struct name "BookInfo" when the prompt was "Tell me about the book X"; disambiguating the prompt fixes it

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: anchpop

Cargo.lock

@@ -1,6 +1,6 @@
 [package]
 name = "tysm"
-version = "0.17.1"
+version = "0.18.0"
 edition = "2021"
 description = "Batteries-included Rust OpenAI Client"
 license = "MIT"
@@ -35,6 +35,7 @@ zstd = "0.13.3"
 log = "0.4.27"
 itertools = "0.14.0"
 dashmap = "6.1.0"
+base64 = "0.22.1"
 
 [dev-dependencies]
 tokio = { version = "1.21.2", features = ["macros", "rt-multi-thread", "full"] }