Release v1.6.1: protect P1 consensus and runtime ownership

[MaxMiksa]

The consensus guard previously accepted a P1 blocker added and immediately checked off during a model cycle. This change rejects that final state, protects existing P1 entries and their descriptions, preserves rejected drafts for human review when possible, restores the pre-cycle consensus and pauses the loop. Agents can still report unresolved blockers; humans resolve them after stopping and completing any interrupted-cycle recovery.

Prepared for v1.6.1. The release preparation commit changes only package.json; the reviewed runtime and documentation remain unchanged.

This PR also delivers the previously reviewed gap fixes:

• Serialize project registry operations and roll back only the failed operation's own entry.
• Keep history and artifact access tied to stable product identity across relocation and same-path replacement.
• Validate macOS service ownership before pause and propagate service-manager failures.
• Restrict ordinary loopback previews to intended public resources and validate request origins and control tokens.
• Align both READMEs with actual runtime boundaries, and document P1 recovery in both languages with the matching Dashboard pause message.

P1 protection compares persisted states at cycle boundaries. It does not guarantee preservation of transient direct edits made while an agent is running, and consensus restoration does not roll back product files or external actions. This PR adds no broad command sandbox, live budget cutoff or business-success gate.

Validation:

• WSL: 48 governance/recovery/fake-engine checks and 4 P1 parser checks passed, with no skips; existing project-governance and loop-integration shell scenarios passed.
• Windows: the same 4 parser checks passed. Frontend Node checks: 19 passed.
• The original four scenarios were compared on old and fixed code. The checked-off P1 case is now rejected, while legitimate unresolved reporting and human-rule protection remain intact. Three new regression tests produce nine expected failing branches against the old code.
• Both language versions have valid local links and synchronized source hashes; showcase sections are unchanged.
• Independent review: no remaining P1/P2 findings; 52 governance/parser tests, 5 shell groups and 9 additional boundary scenarios passed (overlapping coverage, not summed). A discovered logging-failure recovery issue was fixed and independently reproduced as resolved. All 15 changed files match the reviewed hashes.
• Reviewed-code CI 35818189411: all 8 selected jobs passed, including the required CI gate; 3 unchanged example jobs were skipped by change routing. Standard Python: 425 executed/passed, 19 platform/opt-in skips covered by separate macOS launchd (6), real systemd (9), and product-media (43) suites. Browser: 32 passed; frontend Node: 19 passed; both PowerShell editions passed; 197 skill references, none missing. The v1.6.1 version candidate is checked separately in CI 35820049559.
• All model-cycle checks use synthetic engines; no real model or production service was started.