Add generic control kernel for evidence-backed agent workflows

[min9lin9]

Summary

Adds the Stage A generic control-kernel spike approved in #27. Boulder remains the workflow control plane—not the GP domain runtime or agent swarm runtime.

Included

• versioned generic control run-event, evidence-manifest, policy, evaluation, and decision-seal contracts
• mandatory exact local audit recording before evaluation or sealing
• evidence cutoff and SHA-256 manifest/policy binding
• fail-closed hard-failure and promotion-metric evaluation
• failed or blocked tool calls prevent promotion and sealing
• recommendation seals can be created and verified only for eligible runs
• seal hashes are checked before persistence and before idempotent reuse
• contained .boulder/control-kernel/** storage with traversal, symlink, and hardlink checks
• synthetic GP screening fixture covering eligible, hard-failure, metric-failure, incomplete-tool, tampering, and audit-gate paths
• CLI commands and tests for record, evaluate, seal, and verify-seal flows

Safety boundaries

• local-only control artifacts; no provider launch or external transport
• no live investment execution or mutation of deal/fund systems
• SHA-256 seals are tamper-evident bindings, not production identity, quorum, or KMS signatures
• production authority events, append-only storage, long-running state machines, and sandboxed executors remain follow-on work

Validation

• strict TypeScript check: pass
• full Bun test suite: pass
• focused control-kernel unit and CLI tests: pass
• build and package dry run: pass
• package inventory: 191 unique files / 192 packed entries
• official CI workflow on final head fe5c31dc: pass
• CodeQL Security workflow on final head fe5c31dc: pass
• PR is mergeable; it remains draft for human review

Additional deterministic fixes

CI diagnostics exposed two existing environment-sensitive contracts and one stale candidate-package baseline:

• replay and service-readiness fixture discovery now sorts readdir output before constructing evidence strings
• checked published 0.1.16 release evidence is validated against its own frozen manifest, while the current candidate package is validated against the 192-file working-tree inventory
• pack golden comparison still enforces file order, file names, duplicates, package identity, and total count, while normalizing nonsemantic byte-size fields
• the candidate pack dry-run golden baseline was refreshed without modifying the published release evidence manifest

The temporary diagnostic workflow used to isolate the failures was removed before final validation.

Closes #27