Add managed lifecycle and persistent daemon

[wolfiesch]

Why

Chrome owns the lifetime of a Native Messaging process, so service-worker suspension, extension reloads, and native-port churn currently tear down Core IPC, runtime state, and in-flight scheduling. The release lifecycle also needs exact managed update, rollback, uninstall, pruning, and crash recovery before the runtime can be treated as durable infrastructure.

What changed

• installs a persistent per-user agenttab-host daemon plus a small reconnectable agenttab-native relay
• scopes relay traffic and disconnect cleanup to monotonically increasing connection generations
• packages, signs, and verifies both executables on every supported release target
• adds schema-v2 hashed activation receipts with exact file, client-config, Windows registry, and daemon-service ownership
• adds transactional install/update/rollback/uninstall/prune with durable intent/commit recovery and compare-and-swap restoration
• serializes installers across processes with canonical state-directory locks and stable process identity
• rejects symlink targets and unsupported hard-link filesystems before target mutation
• writes and activates per-user launchd, systemd, or limited Scheduled Task definitions without elevation
• restarts the restored service on rollback and disables it during managed uninstall
• retains bounded on-demand daemon startup when the platform service manager is unavailable
• expands agenttab doctor to distinguish installation, IPC, protocol, host, and extension failures

Review decisions

• Persistent user daemon: avoids Core churn without a privileged system service or additional approval prompts.
• Best-effort service activation: a missing or unavailable service manager does not block install or recovery; the native relay remains the fallback.
• Exact ownership, not broad cleanup: rollback and uninstall restore only resources proven by receipts and preserve later user edits.
• One-way durable commit boundary: recovery rolls back only before the commit marker; committed transactions are never heuristically reversed.
• Windows guarantee: process-crash atomicity is covered. Sudden power-loss namespace durability is explicitly not claimed because portable Node does not expose the required Windows directory barrier.
• Unsupported filesystems: hard-link capability is preflighted before mutation; unsupported exFAT/SMB-style environments fail early rather than weakening recovery.

Validation

Exact-head CI run #175 passed:

• macOS architecture/runtime gates, including Rust format, test, and release build
• Linux IPC primitives and complete Rust runtime-core tests
• Windows IPC primitives and complete Rust runtime-core tests
• Ubuntu and Windows adapter/transactional-installer portability
• 76 hosted installer tests, workspace typechecks/builds, protocol/identity/forbidden-surface gates
• local focused installer/lifecycle/service/transaction suite: 68 passed
• local packaging and release-archive tests: 5 passed
• release workflow YAML parse: passed

Why draft

This is a substantial operational architecture change. The low-friction fallback is deliberate, but the persistent-process model, lifecycle semantics, platform service behavior, and explicit Windows durability boundary deserve maintainer review before merge.