R24: witnessed replay rollback and tamper protection

[bitmaster162]

R24 — witnessed replay rollback / tamper protection

Current base: 5a72da3
Current head: b2e32bf
Current tree: 3edc5595bc117d5e04db38bcf4378c713a979a1a
R24 patch SHA-256 vs current master: DF4E64C8621D3E80855CD2017B12DB161FA400864270A7444664871F75AADE4A
R24 patch bytes: 65176

This head explicitly merges current master so the two parallel Remote Commander lines are preserved:

• R23: governed Remote Commander MCP R1 #197 merged as c184ff7
• R23: Secure MCP Tunnel transport R2 #198 merged as 5a72da3
• file overlap with R24: 0
• R24 merge-head parents:
  1. b79df3d
  2. 5a72da3

What R24 changes

• external append-only witness protocol for R23 multi-host replay state
• detects DB rollback/snapshot restore, journal/claim/schema tamper, witness rollback, truncated/reordered history, direct unwitnessed claims
• recovers PostgreSQL materialization from authoritative witnessed history
• preserves R23-compatible claim receipts and R17 approval semantics
• separate witnessed multi-host production entrypoint; no silent fallback to R23/SQLite/memory
• pre-existing R23 claims require explicit migration
• no merge/deploy/runtime/trading/capital authority added

Previously found blocker and fix

Initial adversarial review found external witness I/O under PostgreSQL FOR UPDATE.
Fixed before freeze: DB snapshot/audit -> release DB transaction/row lock -> external witness CAS/read -> reacquire DB -> require unchanged snapshot -> apply witnessed suffix.

Local evidence on synchronized tree

• targeted R17-R24: 54 passed
• release hardening: 10/10 PASS
• repeated full repo: 2280 passed, 2 skipped, 19 subtests
• isolated wheel-only: 2148 passed, 112 skipped, 19 subtests
• source_root_excluded=true
• wheel SHA-256: 8EC0AA04D64EC959CF9D6B4FFCBD59C5193761134AF589DBC19218A889DE4EFF
• compile + git diff --check: PASS

One earlier concurrent local full run reported a Windows loopback socket abort in
tests/test_company_twin.py::test_explorer_is_loopback_only_and_rejects_mutation_routes
(WinError 10053). It is not in the R24 diff. A clean current-master worktree at
5a72da3 reproduced the same failure on iteration 3, proving it is a baseline/local-Windows
flake rather than an R24 regression. A subsequent full R24 run completed 2280/2280 passed.
Natural GitHub CI is the authoritative cross-environment gate.

Qualification boundary

Protocol/core evidence only, not live production qualification.
Production claims still require a real independent append-only witness, shared PostgreSQL,

=2 execution hosts/process domains, network-partition/crash injection, real DB snapshot
restore, bounded witness transport/timeouts/HA/telemetry, and operator recovery/rotation.

Final synchronized exact-head evidence

Current synchronized review bundle SHA-256:
FF106BAA7AAF7D69BE3F6601DC83E288293B3700E91E8E754ED9071909DD8AB1

Fresh synchronized reviewers:

• Hermes: PASS; blockers=0, high=0
• Claude Sonnet 5: PASS; blockers=0, high=0
  • medium hardening notes: string-grep authority smoke-test is weak; narrow read-after-audit TOCTOU window is detected on next synchronize; prior serialization/deadlock retry note remains availability hardening
  • none is a replay bypass; qualification remains explicitly not production-ready

Natural CI on synchronized head b2e32bf:

• P0 Unified Shadow Continuity: SUCCESS
• CodeQL: SUCCESS
• Ubuntu review-gates: SUCCESS
• Windows review-gates: SUCCESS
• Windows full editable-install pytest, wheel-only, governance regression, hardening, Ed25519 verification, receipts: SUCCESS

Baseline-local Windows loopback flake evidence:

• tests/test_company_twin.py::test_explorer_is_loopback_only_and_rejects_mutation_routes can intermittently abort with WinError 10053
• reproduced on clean current master 5a72da3 without R24 on iteration 3
• therefore not attributed to R24
• natural GitHub Ubuntu + Windows review-gates both passed the exact synchronized head