diff --git a/devlog/_plan/260920_meaning_preservation_batch/010_phase2.md b/devlog/_plan/260920_meaning_preservation_batch/010_phase2.md
new file mode 100644
index 00000000000..a6a6a2ff041
--- /dev/null
+++ b/devlog/_plan/260920_meaning_preservation_batch/010_phase2.md
@@ -0,0 +1,84 @@
+# Phase 2 — the ten consolidation bundles
+
+Status: OPEN. Phase 1 (items 1-6) runs in lanes A and B under [000_plan.md](000_plan.md). This file
+opens items 7-16 of the post-2.60.0 assessment. These are **not ten new pull requests**. Each bundle
+is a unit of existing issues and pull requests to reuse, with only the shared part reviewed
+together.
+
+## Delivery topology is unchanged
+
+One branch, ordered commits, one pull request to `dev` per lane. No native stack, no child pull
+request chain. Carried contributor work needs a `Co-authored-by` trailer in a branch commit;
+superseded pull requests are closed by the coordinator only after the lane lands.
+
+## Why these groupings and not one bundle per lane
+
+Bundles 7 and 14 both want the same substrate. Item 7 divides a failure into pre-header,
+headers-only, protocol prelude, semantic output, side effect and terminal, and decides resend
+permission per stage. Item 14 wants a logical request to attempt to physical send to terminal
+record with one cause dictionary. Two lanes defining that separately would typecheck on each branch
+and contradict each other in the merge — the exact class that blocked 2.60.0. They stay in one lane.
+
+Bundles 8 and 9 are the same question asked twice: an observation attributed to an account or
+credential generation must not survive its replacement. A refusal learned from the previous account
+and a warm cache binding dropped on a threshold hint are the same attribution defect at different
+layers.
+
+Bundle 13 consumes the request-scoped route decision that lane B is building for #5087. Starting it
+before lane B lands would fork that authority, so it is scheduled after.
+
+## Lanes
+
+| Lane | Bundles | Existing items to reuse |
+| --- | --- | --- |
+| C | 7 retry stage table, 14 one event model | #4942, #4989, #5245, #2366, #3748, #3983, #5063; issues #4191, #5180 |
+| D | 8 account and credential generation, 9 cache affinity and diagnostics | #5214, #5145, #5229, #5209, #4793; issues #3375, #5178, #3433, #3765 |
+| E | 10 Devin output budget, 11 adapter queue memory, 12 per-key permission | #5189, #5182; issues #5190, #5049 |
+| F (after B) | 13 per-provider egress, 15 CodeBuddy and native wire | #3901, #5148, #5147, #5188; issues #2894, #5146, #5097, #5096 |
+| G (after B) | 16 onboarding, update and screen consolidation | #5016, #4560, #5068; issues #2811, #5215, #5216 |
+
+## Ownership boundaries between concurrent lanes
+
+These exist because the lanes share a checkout-independent surface and would otherwise collide.
+
+- Lane C owns send accounting: `sendCount`, request-wide send budget and the stage and cause
+  vocabulary. Lanes D and E consume it and do not redefine it.
+- Lane D owns #4793 and every per-model cache view. Lane C derives cache projections from the
+  recorder without editing that surface.
+- Lane E owns the adapter event queue budget and the Devin and coding-agent limits. It does not
+  touch retry classification.
+- A collision that cannot be resolved inside these boundaries goes to the coordinator rather than
+  being settled unilaterally in one branch.
+
+## Acceptance that is easy to fake and must not be
+
+Each bundle has a completion condition that a passing request does not demonstrate.
+
+- 7: an uncertain resend after output or a side effect is never automatically permitted; 429, quota,
+  policy refusal and ciphertext refusal stay distinguishable; the provider's stated reason and the
+  actual send count agree, with no duplicated parent and child counter.
+- 8: a refusal or capability observation from a replaced account does not transfer; a cancelled
+  request's late refresh does not overwrite another request's binding; false, unknown and absent
+  stay distinct.
+- 9: passing a threshold alone does not drop a warm binding, while real exhaustion does; input
+  change, account change and transformation change are distinguishable; a prefix fingerprint never
+  becomes a public or durable correlation key.
+- 10: with the caller omitting a limit, the configured effective output cap reaches the wire, and an
+  explicit small cap survives; the history ceiling keeps its own meaning.
+- 11: slow consumers, one large event, accumulated coalescing and a cancel race all stay bounded
+  with no unreleased counter, and a normal long stream is not capped by total length.
+- 12: an alias, combo child, fallback or compact route cannot reach a forbidden model or provider;
+  filtering `/models` is not completion; an inference key never gains management authority.
+
+## Release shape
+
+The first stabilization release carries items 1-6 plus only the small, reproduced fixes from 8, 10
+and 11. Items 7, 9, 12, 13 and 14 form the second group. Items 15 and 16 are the optional extension
+and do not precede the fidelity work. A new large control plane, a full manager rewrite and a
+multi-tenant conversion stay out of this stabilization and are not closed as unwanted.
+
+## Execution constraints
+
+Unchanged from phase 1: no local suites, individual tests, typecheck, build, install or live `ocx`
+execution; verification is static source review plus exact-head hosted CI; pushes use
+`--no-verify`; only the coordinator merges and closes issues.