Record model parameters (max tokens, thinking level) on generations

[wierdbytes]

Problem

pi-observability-plugin does not record any model parameters on generations, so Langfuse cannot show how a model was called — only which model.

Verified in src/index.ts (v0.1.2):

• the generation created in before_provider_request gets model: ctx.model?.id and metadata.provider, nothing else — modelParameters is never set
• rg "modelParameters|thinkingLevel|maxTokens" over src/ returns 0 matches
• stop_reason is recorded on message_end, but the cap that produced a stop_reason: length is not

This hurts when:

• comparing runs of the same prompt at different thinking levels (/thinking high vs low) — the traces are indistinguishable
• diagnosing truncated outputs — you see stop_reason: length but not the max_tokens that was in effect
• reviewing traces later, once the pi settings that produced them have changed

Both values are already on the extension context, and pi resolves every request from the same two fields:

• max output tokens come from model.maxTokens (clamped to the remaining context); the coding agent never overrides it per request, so ctx.model.maxTokens is what goes out
• reasoning comes from the thinking level, or nothing when it is off; each provider maps that level to its own effort/budget field
• temperature and other sampling params are only sent when a caller sets them explicitly, which the coding agent does not

Expected

Every generation carries modelParameters with max_tokens (from ctx.model.maxTokens) and, for reasoning models with thinking on, thinking_level (from ctx.thinkingLevel), so the Langfuse UI shows them next to the model name.

Suggested implementation

1. A pure extractModelParameters(model, thinkingLevel) returning Record<string, string | number> | undefined, with unit tests for the reasoning / non-reasoning / off cases. pi clamps the level to off for non-reasoning models before sending, so a stale level on the context must not produce a misleading chip.
2. In before_provider_request, pass modelParameters: extractModelParameters(ctx.model, ctx.thinkingLevel) to startObservation. @langfuse/tracing drops undefined attributes, so no conditional spread is needed.
3. Integration test on the sandbox model (maxTokens: 8192, reasoning: false) → { "max_tokens": 8192 }; with reasoning: true and defaultThinkingLevel: "high" in settings.json → { "max_tokens": 8192, "thinking_level": "high" }.