Batch 2 — provider correctness / API drift (absorb upstream PRs)

[wishborn]

Second wave of upstream prism-php/prism PRs to review + merge (see .ai/plans/prism-takeover.md in the envelope for full triage).

• fix(gemini): switch structured output to response_json_schema prism-php/prism#975 / issue Gemini: Switch structured output from response_schema to response_json_schema to support full JSON Schema prism-php/prism#974 — Gemini structured output → response_json_schema
• feat(gemini): support mixing provider tools with custom function tools prism-php/prism#976 / issue Gemini: Support mixing provider tools with custom function tools prism-php/prism#972 — Gemini: mix provider tools with custom function tools
• fix(Gemini): Add Usage to StepFinishEvent prism-php/prism#1000 — Gemini: add Usage to StepFinishEvent
• fix(gemini): normalize array-type nullable notation in SchemaMap prop… prism-php/prism#997 — Gemini: normalize array-type nullable notation in SchemaMap
• feat(gemini): add topK generation parameter support prism-php/prism#961 — Gemini: topK parameter
• feat(gemini): add service_tier support for Flex Inference prism-php/prism#1002 — Gemini: service_tier (Flex Inference)
• feat(gemini): extracting rate limit violations and retry-after information  prism-php/prism#1008 — Gemini: rate-limit violations + retry-after extraction
• fix(deepseek): preserve reasoning_content in thinking mode for tool calls prism-php/prism#1020 — DeepSeek: preserve reasoning_content in thinking-mode tool calls
• fix(deepseek): extract prompt_cache_hit_tokens and reasoning_tokens from usage prism-php/prism#1021 — DeepSeek: prompt_cache_hit_tokens + reasoning_tokens usage
• fix(openrouter): support monitoring cache tokens / reasoning tokens on OpenRouter response prism-php/prism#992 / feat(openrouter): extract reasoning tokens and reasoning_details from responses. prism-php/prism#954 / feat(openrouter): support cacheTtl provider option for extended cache TTL prism-php/prism#993 — OpenRouter: cache + reasoning token monitoring, cacheTtl
• Add tool call support to Mistral structured handler prism-php/prism#965 / issue Mistral Structured handler ignores tools in API requests prism-php/prism#963 — Mistral structured handler ignores tools
• feat(mistral): add reasoning_effort provider option prism-php/prism#980 — Mistral: reasoning_effort option
• feat(audio): Allow diarization with Mistral prism-php/prism#995 — Mistral: audio diarization
• feat(zai): add streaming support prism-php/prism#949 — Z.ai: streaming support
• feat(OpenAI): Enable structured output support for OpenAI gpt-5.4 mini and nano models prism-php/prism#999 — OpenAI: structured output for gpt-5.4 mini/nano (verify vs prefix resolver from v0.101.0)
• fix(prism): avoid requiring full laravel framework prism-php/prism#1025 — avoid requiring full laravel/framework

Un-PR'd issues in scope: upstream prism-php#1019, prism-php#990, prism-php#966, prism-php#948, prism-php#940, prism-php#1017, prism-php#979, prism-php#916.

[wishborn]

Done in v0.103.0 — 16 of the 18 PRs absorbed. Exceptions: upstream prism-php#999 superseded by the prefix-based StructuredModeResolver (gpt-5 prefix covers gpt-5.4 models); upstream prism-php#1025 deferred to #3 (illuminate/* split would blind Dependabot to laravel/framework advisories; component list unverifiable under the current test setup — needs an audit + composer-require-checker CI job first).