feat: emit OTel GenAI token and model attributes on LLM spans

[johndemic]

Problem

LLM spans carry token counts and the model only under OpenInference keys
(llm.token_count.*, llm.model_name). Backends that read the OpenTelemetry
GenAI conventions identify the span as GenAI — gen_ai.provider.name is
already set — but then find no usage or model data on it, so they cannot
price the call.

Change

Mirror the existing values onto the GenAI keys. Nothing is renamed or removed,
so OpenInference consumers (Arize/Phoenix) are unaffected:

gen_ai.request.model
gen_ai.response.model
gen_ai.usage.input_tokens
gen_ai.usage.output_tokens
gen_ai.usage.cache_read_input_tokens
gen_ai.usage.cache_creation_input_tokens

Four come from @opentelemetry/semantic-conventions (already a dependency).

Why the cache keys are string literals

semconv spells these gen_ai.usage.cache_read.input_tokens (dotted). Anthropic's
API and the GenAI consumers this targets read the underscored form, so the
constants would emit attributes nothing reads. The new tests assert the literal
strings for the same reason — a constant renamed upstream should fail the suite
rather than silently change the wire format.

Why cache tokens matter here

Not just completeness. A coding agent against a prompt-cached Anthropic model
routinely reports thousands of cache-creation tokens against single-digit input
tokens — one real session showed cache_write=16739 against prompt=3. A
consumer mapping only input/output understates cost by orders of magnitude.

Verification

Tested against a live GenAI backend with a control, 14 seconds apart, identical
token counts — the only difference being attribute naming:

span shape	model resolved	tokens	cost computed
patched (gen_ai.*)	yes	3 in / 5 out / 16739 cache-create	$0.06285525
unpatched (OpenInference only)	unknown	0 / 0	$0.00

The backend derived the cost from token counts alone (no cost attribute was
sent) and its figure matched this plugin's own llm.cost.total to eight
decimal places.

Tests

Three added to tests/handlers/spans.test.ts; all fail against unpatched source.

baseline main: 326 pass
this branch:   329 pass, 0 fail

bun run lint, bun run check:jsdoc-coverage, bun run typecheck all pass.

End-to-end verification

Since opening this, I ran the patched plugin against a live GenAI backend and
captured the wire payload through a logging proxy.

The span the plugin emits:

SPAN opencode.llm  kind=CLIENT  scope=com.opencode
  gen_ai.provider.name                     = anthropic
  gen_ai.request.model                     = claude-sonnet-4-6
  gen_ai.response.model                    = claude-sonnet-4-6
  gen_ai.usage.input_tokens                = 3
  gen_ai.usage.output_tokens               = 5
  gen_ai.usage.cache_read_input_tokens     = 0
  gen_ai.usage.cache_creation_input_tokens = 16919
  llm.cost.total                           = 0.06353025      <- unchanged
  llm.token_count.prompt                   = 3               <- unchanged
  llm.token_count.prompt_details.cache_write = 16919         <- unchanged

The resulting record on the backend:

model                    claude-sonnet-4-6
provider                 anthropic
inputTokenCount          3
outputTokenCount         5
cacheReadTokenCount      0
cacheCreationTokenCount  16919
totalCost                $0.06353025

Every field round-trips, cache tokens included. The backend derived the cost
from token counts alone — no cost attribute is read from the span — and its
figure matched this plugin's own llm.cost.total to eight decimal places.

Without the patch the same traffic records as model: unknown, 0 tokens, $0,
matching the control in the table above.