feat(telemetry): complete metrics instrumentation with cost tracking and per-agent context

[l33t0]

Summary

Continues from #35 which laid the foundation (registry, metrics server, 8 initial metrics, feature gate). This PR resolves every known limitation called out in that PR and adds the operational visibility metrics needed for cost control and agent monitoring.

What's new:

• Metrics server is now wired — start_metrics_server() is called in main.rs, so the /metrics endpoint actually starts
• Per-agent LLM labels — agent_id and tier are no longer hardcoded to "unknown"; SpacebotModel carries process context via .with_context(), wired at all 7 call sites (channel, branch, worker, compactor, ingestion, cortex, cortex_chat)
• Token usage tracking — spacebot_llm_tokens_total counter with direction label (input/output/cached_input)
• Cost estimation — spacebot_llm_estimated_cost_dollars counter with a static pricing table (src/llm/pricing.rs) covering Claude 4/3.5/3, GPT-4o, o-series, Gemini, and DeepSeek families
• Worker/branch visibility — spacebot_active_branches gauge, spacebot_worker_duration_seconds histogram
• Error classification — spacebot_process_errors_total counter with error_type labels (rate_limit, timeout, context_overflow, provider_error, other)
• Memory audit trail — spacebot_memory_updates_total counter tracking save/delete/forget operations; memory_entry_count gauge now wired in MemoryStore
• LLM histogram fix — buckets extended from [0.1 … 10s] to [0.1 … 120s] to capture retry/fallback latency
• Docs — new metrics page for docs.spacebot.sh, updated METRICS.md and docs/metrics.md with full 14-metric inventory, cardinality estimates, and PromQL examples

New metrics

Metric	Type	Labels
spacebot_llm_tokens_total	Counter	agent_id, model, tier, direction
spacebot_llm_estimated_cost_dollars	Counter	agent_id, model, tier
spacebot_active_branches	Gauge	agent_id
spacebot_worker_duration_seconds	Histogram	agent_id, worker_type
spacebot_process_errors_total	Counter	agent_id, process_type, error_type
spacebot_memory_updates_total	Counter	agent_id, operation

Area	Files
Metrics infra	src/telemetry/registry.rs, src/main.rs
LLM instrumentation	src/llm/model.rs, src/llm/pricing.rs (new), src/llm.rs
Agent context wiring	src/agent/{channel,branch,worker,compactor,cortex,cortex_chat,ingestion}.rs
Memory instrumentation	src/memory/store.rs, src/tools/memory_save.rs, src/tools/memory_delete.rs
Hook cleanup	src/hooks/spacebot.rs
Documentation	METRICS.md, docs/metrics.md, docs/content/docs/(deployment)/metrics.mdx (new), meta.json

Test plan

• cargo build --features metrics — compiles (19 pre-existing warnings)
• cargo build (without feature) — compiles (19 pre-existing warnings, no metric code included)
• cargo test --lib --bins — 96 passed, 0 failed
• All #[cfg(feature = "metrics")] gates consistent — no crate::telemetry reference without a gate
• Manual: start with metrics.enabled = true, curl localhost:9090/metrics returns all 14 metrics
• Manual: verify cost counter increments after LLM calls

[l33t0]

Waiting for some time for my local setup to collect enough metrics to create a few Grafana dashboards to visualise the metrics generated here.