feat(runtime): token usage + execution duration emission (closes #87, FWS-3)

[initializ-mk]

Summary

• Emits per-LLM-call token counts (input_tokens / output_tokens — OTel-aligned naming), model, provider, duration_ms, and request_id on every llm_call audit event. Captured directly from provider response metadata across all four providers (Anthropic, OpenAI, Ollama via OpenAI-compatible, OpenAI Responses).
• Emits per-invocation totals as A2A response headers (X-Forge-Tokens-In, X-Forge-Tokens-Out, X-Forge-Duration-Ms, X-Forge-Model, X-Forge-Provider) so orchestrators can enforce cost ceilings inline during parallel workflow execution without subscribing to the audit stream.
• Emits a new invocation_complete audit event with wall-clock duration + aggregated token totals at every A2A request boundary.
• tokens_unavailable=true flag distinguishes "provider did not report usage" (some self-hosted Ollama setups) from "you used zero tokens" so downstream billing doesn't undercount.
• Tool execution events gain duration_ms plus structured arg-shape metadata (args_size, result_size). Raw arg values are deliberately not emitted — that's FWS-8's payload-stripping concern.

Pre-work inventory (per issue body)

Confirmed Architecture A before coding:

Inventory check	Result
Client interface	✓ forge-core/llm/client.go
Normalized response type	✓ ChatResponse with Usage UsageInfo already present
Polymorphic runtime call site	✓ forge-core/runtime/loop.go:245 calls e.client.Chat on llm.Client
Anthropic / OpenAI / OpenAI Responses populate Usage	✓
Ollama populates Usage	Wraps OpenAIClient → handled at audit-emit site via tokens_unavailable

Decision-tree Row 1 → original S (3–5 days) estimate held.

Architectural notes

• Shared call-site instrumentation. AuditLogger.EmitLLMCall is the single capture point for token/duration/model/provider/request_id. The OTel tracing initiative (FORGE_OTEL_TRACING.md) can hook into the same point to populate gen_ai.usage.* span attributes without re-doing per-provider extraction. Same data, captured once, fanned out to multiple emission targets with independent failure domains.
• Field-name alignment with OTel GenAI semconv. Audit emits input_tokens / output_tokens (matching gen_ai.usage.input_tokens / gen_ai.usage.output_tokens). Aligned once at FWS-3, then Forge's audit schema stays Forge-owned and shouldn't churn with upstream OTel renames — consumers correlate via the trace_id/span_id cross-link the OTel work adds later.
• Schema additivity guarantee. All new fields are *int / *int64 + omitempty, so pre-FWS-3 audit consumers parsing session_start / session_end / etc. see byte-identical JSON shape.
• No cost calculation in Forge. Forge emits token counts; the platform applies price tables. Price tables change frequently and shouldn't require agent redeploys.
• A2A headers are the orchestration channel, not the observability channel. They populate regardless of OTel-tracing state.

Wiring

Layer	File
OTel-aligned UsageInfo field names	forge-core/llm/types.go + 4 providers
AuditEvent extension + EmitLLMCall / EmitToolExec / EmitInvocationComplete	forge-core/runtime/audit.go
LLM call timing + provider/model on HookContext	forge-core/runtime/hooks.go + loop.go
Tool-exec timing + arg-shape metadata	loop.go + audit hook in runner.go
Per-invocation LLMUsageAccumulator (thread-safe)	forge-core/runtime/usage_accumulator.go (new)
invocation_complete emission + X-Forge-* headers	forge-cli/runtime/runner.go + forge_usage_headers.go (new)
JSON-RPC tasks/send simplified to delegate to executeTask	forge-cli/runtime/runner.go (~120 lines deleted)

Tests

• forge-core/runtime/audit_llm_test.go — 6 tests: full usage, tokens_unavailable Ollama path, cancelled → llm_call_cancelled, OTel naming check, backward-compat omission for non-LLM events, tool_exec + invocation_complete shape
• forge-core/runtime/usage_accumulator_test.go — 8 tests including a 500-call concurrent-add race regression
• forge-cli/runtime/forge_usage_headers_test.go — 3 tests: full stamping, short-circuited invocation, missing model/provider omission
• forge-core/llm/providers/usage_extraction_test.go — Anthropic / OpenAI / Ollama-no-usage wire-shape tests

Docs

• docs/security/audit-logging.md — new event-types rows (llm_call_cancelled, invocation_complete), expanded llm_call description, new "Token usage and execution duration" section with field table + header table + design notes
• CHANGELOG.md — Unreleased entry above the FWS-1 entry, with the internal UsageInfo rename called out

Test plan

• go test -race -count=1 ./forge-core/... ./forge-cli/runtime/... ./forge-cli/server/... — all 28 packages pass
• golangci-lint run across forge-core/... + forge-cli/... — 0 issues
• gofmt -l clean
• CI green on push

Out of scope (deliberately)

• True streaming llm_call_cancelled emission — the event constant and EmitLLMCall(args.Cancelled) path exist, but ExecuteStream currently wraps non-streaming Chat so the path doesn't fire today. Ready for whenever Forge adopts true client-side streaming.
• Embedding call audit events — embedder.go already uses UsageInfo (now OTel-aligned); per-call audit emission for embeddings is a follow-up that mirrors the llm_call pattern.
• Cost calculation. By design.

Closes #87.