feat(muse-spark): add Meta Muse Spark dashboard template

[gkarthi-signoz]

Adds a dashboard template for Meta Muse Spark accessed through the Meta Model API.

Where the telemetry comes from

Muse Spark is reached over an OpenAI-compatible endpoint (https://api.meta.ai/v1), so the OpenAI SDK plus opentelemetry-instrumentation-openai-v2 traces it with no provider-specific library. That instrumentation alone is not enough for a cost or latency board, for three reasons found while building this:

• Reasoning tokens are never recorded. The extractor never reads completion_tokens_details, and reasoning is 90 to 98 percent of output tokens on every Muse Spark version.
• Cached input tokens are dropped. cache_read_input_tokens is declared and read in util-genai but never assigned, so cost computed from spans overstates real spend by 3 to 5 times on any workload that reuses a prompt.
• Time to first chunk is defined but never emitted, and on this model family it is most of the wall clock on a streaming call.

So the panels read an enriched wrapper span that records those three values from the response object, alongside the stock attributes.

Span shape

span	kind	provider	carries
muse.chat <model>	Internal	meta	usage, reasoning tokens, cache reads, cost, time to first chunk, finish reasons
chat <model>	Client	openai	stock attributes only

Load-bearing query scoping

Every panel filters gen_ai.provider.name = 'meta'. This does two jobs and should not be removed:

1. It separates Muse Spark from other providers reached through the same OpenAI-compatible SDK, which otherwise all report gen_ai.provider.name = openai with no server.address to tell them apart.
2. It prevents double counting. The wrapper span and the SDK client span both carry gen_ai.request.model, so a request count grouped on model alone reports twice the real traffic.

The $model variable is a fixed list rather than a dynamic one, because a dynamic variable populates from every gen_ai.request.model in the workspace and will offer other providers' models on a Muse Spark board.

Panels

Section	Panels
Overview	Total Cost, Requests, Error Rate, Prompt Cache Hit Rate
Cost and Tokens	Cost Over Time by Model, Cost Share by Model, Token Usage by Type, Per-Model Usage and Cost
Latency	Call Duration (p50 / p95 / p99), Time to First Chunk (p50 / p95), Slowest Calls
Reliability	Requests Over Time by Model, Errors Over Time by Model, Finish Reasons, Recent Errors

Notes for review

• The README links https://signoz.io/docs/muse-spark-monitoring/, which does not exist yet. The slug is settled and the guide is in progress.
• Errors cannot be broken down by type. The instrumentation sets no error.type attribute, so a 401, 404 and 400 are indistinguishable except by parsing the span status message. The Errors panel groups by model and the Recent Errors table carries the provider error string.
• Series colours are set explicitly from a colourblind-safe palette. The four model ids differ by one or two characters, so relying on automatic colour assignment made adjacent series hard to tell apart.

[github-actions]

📊 Dashboard files added in this PR

• muse-spark/muse-spark-dashboard.json

✅ An issue has been created in SigNoz/signoz.io to track these dashboard changes.