docs: decision memos — z.ai retention lookup + D07 recommendation

docs/DECISIONS_OPEN_2026-10-04.md

@@ -4,7 +4,7 @@ The single list of decisions only the owner can make, with the context each
 one needs. Everything else open is in [NEXT_SESSION.md](../NEXT_SESSION.md).
 When a decision is made, record it here and in the programme row it affects.
 
-## 1. Model-routing data gates (OPEN)
+## 1. Model-routing data gates (OPEN — terms lookup DONE 2026-10-04, awaiting operator assertion)
 
 **The architecture is right; one fact is missing.** The routing layer gates
 candidates on data class before any model choice: `private` data may only
@@ -22,24 +22,58 @@ The routing shadow (live since 2026-10-04) records on every run:
 This is honest: every run today already sends private-repo data to z.ai via
 the gateway — the shadow makes that explicit instead of implicit.
 
-**The open fact:** what the z.ai coding-plan endpoint actually retains
-(their terms/privacy policy — a lookup, not an engineering slice).
+**The lookup (done 2026-10-04, sources cited):**
 
-**The decision, once the fact is known:**
-- If retention is effectively none: set `retention: "none"` in
-  `/srv/ship-config/routing-policy.json` on infra-home (operator assertion
-  with the terms cited in this file), and the shadow selections go clean.
-- If retention is unclear/lasting: keep the refusals recorded (routing `on`
-  stays blocked for private data) and treat hosted routing as
-  internal/public-data only.
+- **z.ai Data Processing Addendum for API Services, §4(b)**
+  (docs.z.ai/legal-agreement/privacy-policy.md, appended DPA): "The Company
+  do not store any of the content the Customer or its End Users provide or
+  generate while using our Services. This includes any texts, or other data
+  you input. This information is processed in real-time to provide the
+  Customer and End Users with the API Service and is not saved on our
+  servers."
+- **z.ai Privacy Policy, §3 purpose table**: the "improve and develop our
+  Services and conduct research … including when we train and improve our
+  models" row lists account/communication/log/usage/device/cookie data —
+  **User Content (prompts) is not among the training-purpose categories**;
+  prompts appear only under "provide … the Services" (contract basis). §5:
+  deleted conversations "will be removed immediately … and automatically
+  deleted from our back-end."
+- **GLM Coding Plan usage policy** (docs.z.ai/devpack/usage-policy.md) and
+  subscription terms: no retention or training carve-out for the coding-plan
+  endpoint; the plan is provisioned on the API platform
+  (z.ai/manage-apikey/* console paths).
 
-## 2. S12 browser harness — D07 (OPEN)
+**Residual ambiguity (why this still needs the operator, not the agent):**
+the DPA is framed for "business and enterprises users … through API
+Services", and the coding plan is an individual subscription riding the API
+platform's endpoint — no document states in words that the *coding-plan
+endpoint specifically* retains nothing. The evidence is strong (API surface
+= no content storage; the consumer policy never trains on prompt content)
+but the seam between the two policies is exactly where a cautious reading
+keeps the refusal.
+
+**Recommended assertion (one line, when accepted):** in
+`/srv/ship-config/routing-policy.json` on infra-home set the
+`api.z.ai`/coding-plan destination to `"retention": "none"`, citing the two
+clauses above, with a re-check trigger on any z.ai privacy-policy date
+bump (current: 2025-09-29). Until asserted, the shadow refusals keep
+recording — routing stays honest either way.
+
+## 2. S12 browser harness — D07 (OPEN — recommendation recorded 2026-10-04)
 
 The last undecided design call (deferred as D07). Options: keep the current
 harness approach and scope remaining S12 browser/interactive work on it, or
 run a short evaluation slice of harness options first and then build. S12
 remains the least-started package; nothing blocks on it.
 
+**Recommendation (agent, for the owner to accept or overrule): keep the
+current harness; skip the evaluation slice.** Rationale: nothing blocks on
+S12, so the only thing the eval buys is optionality; the eval itself costs
+a full session; and any harness shortcoming will announce itself in the
+first S12 interactive milestone, at which point a *targeted* swap is better
+informed than an upfront bake-off. Re-open D07 only if that first milestone
+lands and the harness measurably fights the work.
+
 ## Decided 2026-10-04 (context)
 
 - **S02 batch**: A-only via the coding-plan key ($0 real); B deferred.