fix: phase 3 medium findings (token estimate, seatbelt, pool errors)

[Patel230]

Summary

Three Phase 3 medium findings from the deep audit, all hawk-internal:

1. Token estimate — estimateTokens in selective_rag.go used len(text)/4, which systematically undercounts code-heavy text and drives compaction decisions. Now uses the BPE-based token.CountTokensFast (tok tokenizer) that the streaming path already uses.
2. Seatbelt mach-lookup — the macOS seatbelt profile emitted an unconditional (allow mach-lookup) (every XPC service). Non-TierOff tiers now use a minimal service allowlist (opendirectoryd, notification center, cfprefsd, logger; plus DNS/network-config services only when network is allowed). TierOff keeps legacy behavior.
3. Pool.Run silent failures — parallel.Pool.Run always returned nil even when every task failed. It now returns errors.Join of all failed tasks (task description + error); tasks that fail still don't block others. CLI caller (cmd/chat_commands.go) already renders per-pane failure state and now also receives the aggregate error.

Changes

• internal/engine/selective_rag.go — estimateTokens → token.CountTokensFast
• internal/engine/conversation_adapter_test.go — drop len/4 formula assertion; sanity-range check instead
• internal/sandbox/seatbelt.go — tiered mach-lookup scoping
• internal/sandbox/seatbelt_test.go — new TestGenerateSeatbeltProfile_MachLookupTiered
• internal/multiagent/parallel/parallel.go — aggregate failure error from Run
• internal/multiagent/parallel/parallel_test.go — updated TestErrorHandling, TestContextCancellation

Testing

• go test -race ./... — all packages pass
• golangci-lint — 0 issues; gofmt clean
• GitNexus detect_changes: expected symbols only, no affected execution flows

Out of scope (live in external/ repos, separate PRs): inline tool-call parser (eyrie), PostHog telemetry (trace repo).