docs: align mainline with community reliability signals

README.md

@@ -148,7 +148,8 @@ pending after a reproducible local DNS/TLS endpoint failure; it has not been
 upgraded into product evidence. The deterministic manual-accounting baseline,
 active-receiver checkpoint negative, and 76-second evidence walkthrough are now
 the public product proof. New harness, transport, and generalized protocol work
-remains frozen until the live rerun and three external setup attempts close.
+remains frozen until the live rerun, measured manual/live baseline, and three
+external setup attempts close.
 
 [Read the exact fixture evidence](docs/09-reviews/2026-09-01-m5-2-autonomous-fixture.md) ·
 [Read the real Codex behavior](docs/09-reviews/2026-09-01-m5-2-real-codex-event-pump-behavior.md) ·
@@ -172,11 +173,11 @@ npm ci
 npm run demo
 ```
 
-This deterministic demo creates four isolated sessions and runs an
-implementation → review → fix → review → dependent-task sequence. It exposes
-the event, routing reason, receiver decision, external verification,
-dependency effect, and cleanup state without spending model quota or touching
-your agent sessions.
+This deterministic demo creates four workflow sessions plus one isolated
+active-receiver safety task. It runs an implementation → review → fix → review
+→ dependent-task sequence and exposes the event, routing reason, receiver
+decision, fixture-signed verified disposition, dependency effect, and cleanup
+state without spending model quota or touching your agent sessions.
 
 [Read the attention-router demo guide](docs/06-guides/attention-router-demo.md)
 

README.zh-CN.md

@@ -110,9 +110,10 @@ npm ci
 npm run demo
 ```
 
-这个确定性案例会创建四个隔离 session，并执行“实现 → 评审 → 修复 → 再评审 →
-解锁下游任务”的完整链路。输出会分别展示事件、路由理由、接收方决定、外部验证、
-依赖效果与清理状态；它不消耗模型额度，也不触碰你的真实 Agent session。
+这个确定性案例会创建四个工作流 session，加一个隔离的活跃接收方安全任务，并执行
+“实现 → 评审 → 修复 → 再评审 → 解锁下游任务”的完整链路。输出会分别展示事件、
+路由理由、接收方决定、fixture 签名的验证 disposition、依赖效果与清理状态；它不
+消耗模型额度，也不触碰你的真实 Agent session。
 
 [查看 attention-router 演示指南](docs/06-guides/attention-router-demo.md)
 

ROADMAP.md

@@ -199,6 +199,12 @@ partial integration attempt is promoted to M5.2 evidence.
     output.
   - [ ] M5.3: pass three fresh relevant runs plus the manual baseline,
     irrelevant, stale/unverified, restart, and cleanup matrix.
+- [ ] Add one bounded, read-only correlated handoff state vector after the
+  current live-value gate, without adding a scheduler or transport
+  ([#135](https://github.com/fyaic/threadmesh/issues/135)).
+- [ ] Prove a real busy receiver is not silently steered and preserve admitted
+  context across restart/compaction through a durable attention inbox
+  ([#136](https://github.com/fyaic/threadmesh/issues/136)).
 - [ ] Repeat the loop across Codex and one ACP-compatible harness
   ([#93](https://github.com/fyaic/threadmesh/issues/93)).
 - [x] Publish the bounded inspector and reproducible deterministic evidence
@@ -227,6 +233,9 @@ incorrect dependency unlocks.
 - [ ] Prototype a Cotal transport bridge only after the local loop passes.
 - [ ] Receive one external connector contribution or equivalent clean-room
   integration.
+- [ ] After #91 and #93 each have real evidence, add durable workflow budgets,
+  recursive stop, and an owner-visible circuit breaker
+  ([#137](https://github.com/fyaic/threadmesh/issues/137)).
 
 Exit: an operator outside the maintainer organization completes a useful loop
 without maintainer intervention, and the integration contract is ready for a

docs/07-research/community-signals.md

@@ -1,9 +1,26 @@
 # Community signals around cross-task agent coordination
 
-> Research snapshot: 2026-09-01. GitHub issues are community reports unless an
+> Research snapshot: 2026-09-02. GitHub issues are community reports unless an
 > OpenAI contributor explicitly confirms product status. They are evidence of
 > demand and failure modes, not a stable API contract.
 
+## 2026-09-02 reliability and operator-control update
+
+New reports sharpen three product requirements beyond transport:
+
+| Requirement | Community reports | ThreadMesh decision |
+|---|---|---|
+| Correlated, complete handoff state | Codex [#29886](https://github.com/openai/codex/issues/29886), [#40397](https://github.com/openai/codex/issues/40397), [#41964](https://github.com/openai/codex/issues/41964), [#41972](https://github.com/openai/codex/issues/41972), [#42074](https://github.com/openai/codex/issues/42074) | Track one bounded state vector in [#135](https://github.com/fyaic/threadmesh/issues/135); do not require another model turn to learn delivery state. |
+| Durable, non-interrupting attention with fresh context | Codex [#30499](https://github.com/openai/codex/issues/30499), [#41995](https://github.com/openai/codex/issues/41995), [#42131](https://github.com/openai/codex/issues/42131) | Track real busy-target and context-epoch behavior in [#136](https://github.com/fyaic/threadmesh/issues/136); queue/checkpoint remains the default. |
+| Hard autonomy limits and recursive stop | Codex [#35177](https://github.com/openai/codex/issues/35177), [#40037](https://github.com/openai/codex/issues/40037), [#42032](https://github.com/openai/codex/issues/42032), [#42080](https://github.com/openai/codex/issues/42080), [#42182](https://github.com/openai/codex/issues/42182) | Track workflow-level budgets and circuit breakers in [#137](https://github.com/fyaic/threadmesh/issues/137), but freeze implementation until real Codex and cross-harness value is established. |
+
+The reports validate ThreadMesh's attention/admission thesis, but they do not
+authorize a project orchestrator. Project-level contractor mode, Codex writer
+takeover, UI persistence fixes, model routing, and a new scheduler remain
+upstream concerns or explicit non-goals. ThreadMesh should expose typed
+detect/defer/fail-closed behavior at those boundaries instead of replacing the
+harness.
+
 ## 2026-09-01 delivery-mode update
 
 The most useful new distinction is no longer whether Codex can retain queued
@@ -53,15 +70,17 @@ to:
 
 ### Community and adoption reality
 
-At this snapshot, ThreadMesh has no external stars, forks, watchers, issue
-comments, or independent review records. Before this roadmap update, its two
-open issues were the external M0 review gate
-[#7](https://github.com/fyaic/threadmesh/issues/7) and independent
-harness-author feedback [#79](https://github.com/fyaic/threadmesh/issues/79).
-The new M5 issues [#89](https://github.com/fyaic/threadmesh/issues/89)–
-[#93](https://github.com/fyaic/threadmesh/issues/93) are maintainer planning,
-not community adoption. Repository clone traffic is not treated as adoption
-because CI, dependency bots, and maintainer validation can dominate it.
+At this snapshot, ThreadMesh has one star, zero forks, zero external issue
+comments, and no independent review or setup record. The star is a positive
+discovery hint, not evidence that an operator reached value. Open product and
+review work remains maintainer-defined: external M0 review
+[#7](https://github.com/fyaic/threadmesh/issues/7), independent operator
+feedback [#79](https://github.com/fyaic/threadmesh/issues/79), the Codex and
+cross-harness gates [#91](https://github.com/fyaic/threadmesh/issues/91) and
+[#93](https://github.com/fyaic/threadmesh/issues/93), plus the three grouped
+community backlogs #135–#137. Repository clone traffic is not treated as
+adoption because CI, dependency bots, and maintainer validation can dominate
+it.
 
 This is neither product rejection nor validation: the repository was created
 only eight days earlier, but there is currently no evidence that an external

docs/09-reviews/2026-09-02-mainline-checkpoint.md

@@ -0,0 +1,72 @@
+# 2026-09-02 mainline checkpoint
+
+## Decision
+
+The mainline remains the real Codex implementation → review → same-A fix →
+verification → dependent loop in issue #91. The repository is not currently
+implementation-blocked; it is live-evidence and adoption-blocked. No new
+transport, scheduler, harness, or general orchestration subsystem enters the
+critical path.
+
+## Repository state
+
+- branch: `main`, clean and synchronized with `origin/main`;
+- commit: `0c7165ead39499497827737c17e94e95f0f3286b`;
+- latest product work: real-effects integration #133 and product-proof update
+  #134;
+- public discovery: one star, zero forks, and no external operator report;
+- open pull requests: Dependabot maintenance only.
+
+## Local evidence retained
+
+The following passed on the exact checkpoint commit:
+
+- 384/384 unit and subtests;
+- 55 schema cases and 7 transition cases;
+- 112 Markdown files with zero lint issues;
+- focused autonomous real-effects scenario: 9/9;
+- one-command attention-router demo: `state=passed`;
+- modeled manual lower bound: 9 user actions;
+- ThreadMesh demo path: 1 kickoff, 0 relay actions, 0 status checks;
+- active receiver remains `running` with a pending `checkpoint-offer`, 0 steer,
+  0 interrupt, and 0 native-turn starts;
+- dropped wake hints reconcile 4/4 and cleanup is complete.
+
+The 9-to-1 comparison remains modeled workflow accounting. Elapsed time and
+model tokens are not measured.
+
+## Live-product preflight
+
+Codex CLI 0.145.0 remained authenticated and locally healthy. The Responses
+WebSocket diagnostic timed out, system DNS returned unexpected non-provider
+addresses for `chatgpt.com`, and a certificate-verifying HTTPS probe timed out
+during the SSL connection. No DNS override, TLS bypass, live session creation,
+or real Git-effect gate was attempted.
+
+This is an environment blocker, not positive or negative ThreadMesh product
+evidence. The next live attempt begins only after normal DNS and certificate-
+verified connectivity are restored.
+
+## Ordered next gates
+
+1. One network-valid traversal of the merged real-effects Codex path.
+2. One same-condition manual/ThreadMesh comparison with action, elapsed, usage,
+   interruption, duplicate-delivery, and cleanup measurements.
+3. The complete M5.3 matrix: relevant 3/3, irrelevant, stale/unverified,
+   restart/replay, and failure cleanup.
+4. Close #91 only after its canonical checklist passes.
+5. Observe three independent 15-minute operators for #79.
+6. Prioritize #135 and #136 from observed friction; keep #137 frozen until #91
+   and #93 have real evidence.
+
+## Community backlog decision
+
+Community reports were grouped by user-visible failure mode rather than copied
+one-for-one into the repository:
+
+- #135: correlated handoff receipt and complete attention state;
+- #136: durable non-interrupting attention plus context freshness;
+- #137: owner-visible autonomous-chain budgets and circuit breakers.
+
+These issues deliberately exclude project orchestration, model routing, Codex
+writer takeover, UI bug fixes, hosted control planes, and new event buses.

docs/09-reviews/README.md

@@ -30,6 +30,7 @@ not count as live-product or independent external-verifier evidence.
 
 ## Live attempt audits
 
+- [2026-09-02 mainline and connectivity checkpoint](2026-09-02-mainline-checkpoint.md)
 - [M5.2 real Codex event-pump attempt audit](2026-09-01-m5-2-real-codex-event-pump-attempt-audit.md)
 - [Real Codex autonomous event-pump behavior](2026-09-01-m5-2-real-codex-event-pump-behavior.md)
 

docs/10-planning/mainline-plan.md

@@ -15,7 +15,7 @@ inspector, and an independently repeated implementation/review/fix loop.
 Normative M0 review remains a parallel governance track. Maintainer-authorized
 experiments are explicitly labeled and do not satisfy it.
 
-## Active critical path — 2026-09-01
+## Active critical path — 2026-09-02
 
 M5.2 remains the only implementation critical path. The behavioral question is
 answered by the completed sixth Codex run. The bounded Git effects and existing
@@ -51,18 +51,23 @@ Execute in this order:
 3. Retain the measured manual/live baseline. The executable demo already
    reports a nine-action manual lower bound versus one kickoff, but elapsed time
    and tokens remain explicitly unmeasured.
-4. Observe three independent operators using only the README and 15-minute
+4. Complete the M5.3 closure matrix: three fresh relevant runs, irrelevant,
+   stale/unverified, restart/replay, and exact failure cleanup. Close #91 only
+   after that canonical checklist and the measured baseline pass.
+5. Observe three independent operators using only the README and 15-minute
    challenge; convert friction into product or documentation fixes.
-5. Close #91 only when its original outcome is satisfied, then resume
-   repetition, Kimi parity, and production-hardening evidence.
+6. Use the resulting evidence to prioritize the correlated state vector #135
+   and durable non-interrupting attention inbox #136 before broader adapter
+   work. Keep the autonomous circuit breaker #137 frozen until #91 has a live
+   pass and #93 has a real cross-harness pass.
 
 Mainline guardrail: do not add a new substrate or generalize an existing one
 unless the current live chain demonstrates that it is the blocking condition.
-OS-kill matrices, long-turn heartbeat, a global cross-dispatch chain, new
-harnesses, A2A/Cotal work, hosted operation, and new transport/protocol surface
-are paused. The evidence walkthrough is complete and must now be judged by
-external first-run attempts. Small PRs remain acceptable; scope, not PR size,
-is the constraint.
+OS-kill matrices beyond the canonical restart/replay case, long-turn heartbeat,
+a global cross-dispatch chain, new harnesses, A2A/Cotal work, hosted operation,
+and new transport/protocol surface are paused. The evidence walkthrough is
+complete and must now be judged by external first-run attempts. Small PRs
+remain acceptable; scope, not PR size, is the constraint.
 
 The detailed historical gates remain in the
 [real Codex scenario plan](m5-2-real-codex-scenario.md). The

docs/10-planning/product-mainline-2026-08-28.md

@@ -216,18 +216,23 @@ labels its real-versus-deterministic evidence boundary.
 Outcome: someone outside the maintainer organization integrates an existing
 harness and completes a real loop without maintainer intervention.
 
-1. Turn the existing ACP adapter into the preferred multi-harness gateway and
-   document the registry-compatible path.
-2. Publish an A2A mapping for task, message, artifact, lifecycle, and extension
-   fields; keep receiver admission local to ThreadMesh.
-3. Prototype a Cotal transport bridge only after the core loop works locally.
-4. Recruit three external operators for a 15-minute setup task. Capture
+1. Recruit three external operators for a 15-minute setup task. Capture
    installation outcome, time to first value, failed steps, and whether they
    would use the loop again.
-5. Close [#79](https://github.com/fyaic/threadmesh/issues/79) only with feedback
+2. Close [#79](https://github.com/fyaic/threadmesh/issues/79) only with feedback
    from an independent harness author; seek one external connector pull
    request.
-6. Prepare `0.1` only after the workflow and integration contract survive the
+3. Use live and operator evidence to add one bounded correlated handoff state
+   vector [#135](https://github.com/fyaic/threadmesh/issues/135) and the real
+   busy-target/context-freshness behavior in
+   [#136](https://github.com/fyaic/threadmesh/issues/136).
+4. Turn the existing ACP adapter into the preferred multi-harness gateway and
+   complete the real cross-harness gate
+   [#93](https://github.com/fyaic/threadmesh/issues/93).
+5. Publish an A2A mapping for task, message, artifact, lifecycle, and extension
+   fields; keep receiver admission local to ThreadMesh.
+6. Prototype a Cotal transport bridge only after the core loop works locally.
+7. Prepare `0.1` only after the workflow and integration contract survive the
    external attempts.
 
 ### Later — harden only proven usage
@@ -238,6 +243,9 @@ observed failure modes.
 - claimant-specific mailbox leases and crash recovery;
 - production authentication and remote transport profiles;
 - OS isolation and secret minimization;
+- durable workflow budgets, recursive stop, and an owner-visible circuit
+  breaker [#137](https://github.com/fyaic/threadmesh/issues/137), only after
+  #91 and #93 each have real evidence;
 - hosted or multi-user coordinator only with demonstrated demand;
 - additional wake, steer, or interruption capabilities only when a validated
   workflow requires them.