feat(audit): add durable agent loop tracing

[geminixiang]

Summary

Adds a deployment-owned, metadata-only agent-loop audit subsystem and a dedicated Admin Audit view.

• creates the top-level runId at runtime admission and correlates Sentry, runner, harness, tools, logical model requests, compaction, retry, budgets, Session Dream, and child subagents
• stores immutable typed events plus run/tool/model projections in <state-dir>/audit/audit.sqlite
• keeps SQLite off the agent hot path with a bounded non-throwing queue and one worker-thread writer
• adds WAL/busy handling, exclusive schema migration, owner-only file modes, run-coherent 90-day retention, health/degraded counters, and bounded shutdown
• adds Admin filters for conversation, run ID, tool, status, and time, with keyset run pagination and a backward-pageable complete run timeline
• deliberately stores no prompts, completions, thinking, tool arguments/results, provider payloads/headers, images, file bodies, terminal output, or free-form error messages

Architecture rationale and Codex rollout-trace research are captured in:

• docs/adr/0006-durable-agent-audit-and-diagnostic-traces.md
• docs/research/mikan-agent-audit-trace-architecture.md

How I tested

• npm run lint
• npm run fmt:check
• npm run knip
• npm run build
• npm test — 121 files / 1770 tests
• Manual verification

Manual verification:

• built-distribution worker smoke test against dist/audit/worker.js
• 20 consecutive dual-store concurrent first-boot/WAL startup checks
• architecture TOML/reference/Markdown-link validation
• Admin inline JavaScript syntax compilation in tests
• byte-for-byte verification that the unrelated untracked research file was excluded

Notes for reviewers

• This is the run-centered Phase 1 described in the research: indexed run search plus pageable per-run events. Cross-run event search is intentionally deferred.
• Codex-like raw diagnostic capture is intentionally not included. It needs separate opt-in enablement, retention, authorization, and evidence policy.
• node:sqlite still emits Node's experimental-feature warning on the supported Node releases; no new dependency was added.
• Audit write/serialization/worker failures never change the agent outcome. Loss is surfaced through Audit health and drop/degraded counters.

[cloudflare-workers-and-pages]

Deploying mikan with    Cloudflare Pages

Latest commit:	748166b
Status:	✅  Deploy successful!
Preview URL:	https://e5b2c96e.mikan-1wv.pages.dev
Branch Preview URL:	https://feat-agent-audit-trace.mikan-1wv.pages.dev

View logs