diff --git a/.conductor/registry/workflows/apex-item-dispatch.yaml b/.conductor/registry/workflows/apex-item-dispatch.yaml
index 467db188..a3e3db78 100644
--- a/.conductor/registry/workflows/apex-item-dispatch.yaml
+++ b/.conductor/registry/workflows/apex-item-dispatch.yaml
@@ -146,11 +146,20 @@ output:
   # ── Renegotiation bubble-up (PR #144 — plan-level outputs) ──
   # Only plan-level emits these; the other lifecycles leave them
   # undefined. The `is defined` guards collapse to safe defaults.
+  #
+  # `renegotiation_request` is bubbled up from
+  # PlanExtractRenegotiationFlagResult.RenegotiationRequest which is
+  # `string?` and OMITTED when null. plan-level.yaml's output: block
+  # always emits the field as a string (post-fix at the plan-level
+  # source) but if plan-level itself crashes mid-run, the .output
+  # dict can be partial. Two-level guard is defense-in-depth against
+  # any future plan-level failure mode that leaves outputs incomplete.
+  # Cf. plan-level.yaml output: block for the source-of-truth fix.
   renegotiation_pending: >-
     {%- if plan_level_dispatch is defined -%}{{ (plan_level_dispatch.output.renegotiation_pending | default(false)) | string | lower }}
     {%- else -%}false{%- endif -%}
   renegotiation_request: >-
-    {%- if plan_level_dispatch is defined -%}{{ plan_level_dispatch.output.renegotiation_request | default('') }}
+    {%- if plan_level_dispatch is defined and plan_level_dispatch.output.renegotiation_request is defined -%}{{ plan_level_dispatch.output.renegotiation_request }}
     {%- else -%}{%- endif -%}
   validate_scope_verdict: >-
     {%- if plan_level_dispatch is defined -%}{{ plan_level_dispatch.output.validate_scope_verdict | default('') }}
diff --git a/.conductor/registry/workflows/github-pr.yaml b/.conductor/registry/workflows/github-pr.yaml
index b9e243af..38b4f4d8 100644
--- a/.conductor/registry/workflows/github-pr.yaml
+++ b/.conductor/registry/workflows/github-pr.yaml
@@ -55,8 +55,18 @@ output:
   # Lowercase-coerce so the parent workflow's _maybe_parse_json sees
   # 'true'/'false' rather than the capital 'True'/'False' Jinja produces
   # for Python booleans (see conductor-mechanics M7).
-  merged: "{{ (pr_merger.output.merged | default(false)) | string | lower }}"
-  pr_url: "{{ pr_merger.output.pr_url | default('') }}"
+  #
+  # pr_merger may not run — review_router routes around it on
+  # changes_requested, on rev-cap exhaustion (force-abort), and on
+  # human-gate abort. Per M3 (StrictUndefined), guard `pr_merger is
+  # defined` first; without the guard, accessing pr_merger.output
+  # raises TemplateError on every non-merge terminal.
+  merged: >-
+    {%- if pr_merger is defined -%}{{ (pr_merger.output.merged | default(false)) | string | lower }}
+    {%- else -%}false{%- endif -%}
+  pr_url: >-
+    {%- if pr_merger is defined -%}{{ pr_merger.output.pr_url | default('') }}
+    {%- else -%}{%- endif -%}
 
 agents:
   # ── PR reviewer ────────────────────────────────────────────────────────
diff --git a/.conductor/registry/workflows/implement-pg.yaml b/.conductor/registry/workflows/implement-pg.yaml
index b4ebf456..369e0db9 100644
--- a/.conductor/registry/workflows/implement-pg.yaml
+++ b/.conductor/registry/workflows/implement-pg.yaml
@@ -69,8 +69,18 @@ output:
   # several routes (pr_router action='all_complete', pr_lifecycle merged=
   # false, dependency_gate reassign, user_acceptance abort) reach $end
   # without running scope_closer.
+  #
+  # pr_submit is an LLM agent whose `pr_url` field is NOT marked
+  # `required: true` in its output schema, so the LLM may emit a
+  # well-formed envelope without the field. The `pr_submit is defined`
+  # guard alone is insufficient; we also need to check
+  # `pr_submit.output.pr_url is defined` to avoid StrictUndefined on a
+  # missing dict key. Cf. plan-level.yaml output: block for the same
+  # trap on a verb-side nullable field.
   merged: "{% if scope_closer is defined %}{{ (scope_closer.output.exit_code == 0) | string | lower }}{% else %}false{% endif %}"
-  pr_url: "{% if pr_submit is defined %}{{ pr_submit.output.pr_url | default('') }}{% else %}{% endif %}"
+  pr_url: >-
+    {%- if pr_submit is defined and pr_submit.output.pr_url is defined -%}{{ pr_submit.output.pr_url }}
+    {%- else -%}{%- endif -%}
 
 agents:
   # ── PG router ──────────────────────────────────────────────────────────
diff --git a/.conductor/registry/workflows/plan-level.yaml b/.conductor/registry/workflows/plan-level.yaml
index de4fd6e2..b14cc6fe 100644
--- a/.conductor/registry/workflows/plan-level.yaml
+++ b/.conductor/registry/workflows/plan-level.yaml
@@ -144,20 +144,41 @@ tools:
 # M7 booleans are piped through `| string | lower` so the parent receives
 # a real bool rather than capital "True"/"False".
 #
+# StrictUndefined trap (2026-05-10 dogfood AB#3067, plan-PR #260):
+# `<agent> is defined` only checks the agent variable, NOT the dict keys
+# under `.output`. When a verb's Result model declares `string? Foo`
+# (omitted from JSON when null per WhenWritingNull), accessing
+# `<agent>.output.foo` RAISES TemplateError — `| default('')` does not
+# rescue, because StrictUndefined fires on the missing dict key BEFORE
+# the filter runs. So nullable verb fields require a TWO-LEVEL guard:
+# `<agent> is defined and <agent>.output.<field> is defined`. Required
+# fields (e.g. `flag_present`, `verdict`) are always emitted and only
+# need the agent-level guard. Cf. PlanDeriveAncestorChainResult.cs which
+# uses `[JsonIgnore(Condition = Never)]` to dodge the same trap from the
+# verb side.
+#
 # Design rationale — option (C) in docs/decisions/scope-renegotiation.md
 # (handler design — workflow output bubble-up). Today plan-level has no
 # caller, so these outputs go nowhere; that is intentional. apex-driver
 # (a downstream PR) consumes them.
 output:
+  # FlagPresent is `required bool` in PlanExtractRenegotiationFlagResult
+  # — always emitted, agent-level guard suffices.
   renegotiation_pending: >-
     {% if extract_renegotiation_flag is defined %}{{ (extract_renegotiation_flag.output.flag_present | default(false)) | string | lower }}
     {%- else %}false{% endif %}
+  # RenegotiationRequest is `string?` — OMITTED from JSON when no flag
+  # was present. Two-level guard is mandatory; see the trap note above.
   renegotiation_request: >-
-    {% if extract_renegotiation_flag is defined %}{{ extract_renegotiation_flag.output.renegotiation_request | default('') }}
+    {% if extract_renegotiation_flag is defined and extract_renegotiation_flag.output.renegotiation_request is defined %}{{ extract_renegotiation_flag.output.renegotiation_request }}
     {%- else %}{% endif %}
+  # Verdict is `required string` in PlanValidateScopeResult — always
+  # emitted, agent-level guard suffices.
   validate_scope_verdict: >-
     {% if validate_scope is defined %}{{ validate_scope.output.verdict | default('') }}
     {%- else %}{% endif %}
+  # FilesOutOfScope is non-nullable IReadOnlyList<string> with default
+  # `Array.Empty<string>()` — always emitted, agent-level guard suffices.
   scope_violation_files: >-
     {% if validate_scope is defined %}{{ validate_scope.output.files_out_of_scope | default([]) | tojson }}
     {%- else %}[]{% endif %}