feat: recover unlinked live daemon registry

[schickling-assistant]

Closes #126.

Summary

Adds a fail-closed pty recover-live --metadata <snapshot.json> <name> path for
future daemons whose pathname socket and registry sidecars are accidentally
unlinked while their provider is still alive.

The original daemon watches for a short-lived authenticated request and:

• validates name, daemon PID, generation, strong OS process-start token, and
  the full persisted launch shape;
• refuses conflicting pid/metadata generations and foreign socket inodes;
• closes only its stranded listener, rebinds in the same daemon, and republishes
  the captured metadata plus authoritative launch fields;
• preserves the provider and already-established clients;
• serializes recovery with shutdown and supports retry after a failed bind.

There is no process signal and no provider restart. Unsupported snapshots fail
before publishing a request.

Important applicability boundary

This does not recover daemons that were already running before this code.
Those processes have no request watcher or protocol/start-token marker and
cannot be upgraded in place. The CLI deliberately refuses their snapshots;
they must remain alive and use an existing attachment or transcript fallback.

Reproduction

https://github.com/schickling-repros/2026-07-pty-unlinked-live-daemon

Verification

• npm run typecheck
• npm run build
• recovery/help/completion/disk-format focused tests: 49/49
  • bidirectional established-client continuity and provider PID liveness
  • unsupported snapshot refusal without poking the daemon
  • foreign listener refusal, removal, then successful retry
  • stale request refusal followed by fresh replacement
  • bind-race failure after listener close followed by successful retry
  • nonce-specific concurrent request removal isolation
  • filename/payload nonce binding
  • cross-session cleanup isolation
• black-box eval probe: all 7 green markers

Paired eval: compoundingtech/evals#54 and compoundingtech/evals#55.

Posted on behalf of @schickling

field	value
agent_name	co2-bear
agent_session_id	a078daee-6f98-4916-91a8-d21291407789
agent_tool	Codex CLI
agent_tool_version	0.145.0
agent_runtime	Codex CLI 0.145.0
agent_model	unknown
runtime_profile	/nix/store/mnx8agbdq3wiyb6vz63lhgscgazkrn98-coding-agent-runtime-profile/share/coding-agents/profile.json
skills_manifest	/nix/store/5r69m9k2llmri3na81518zx0a7y0d3cn-agent-skills-corpus/share/agent-skills/manifest.json
worktree	pty/schickling-assistant/2026-07-30-swift-lamarr-46
machine	dev3
tooling_profile	dotfiles@0fb7e03

[myobie]

Closing as superseded by #128, the selected issue #126 implementation.

The two approaches are protocol-incompatible. Exact review found #127 does not provide the accepted capability/HMAC authorization, signed mutable-metadata revision binding, or private root/.recovery identity and permission binding; it also relies on a liveness probe. PR #128 now has independent semantic GREEN at exact head f70f85baa89145136a63cc9909fc2423bd3d1e7d, deterministic stale-snapshot/interrupted-publication/crash-resume/privacy regressions, and green hosted Nix and Test gates.