fix(substrate): TOML scope-capture + dangling physics-deriver bodies + rule paths

Three independent manifest fixes from Opus TOML + Sonnet TOML + Sonnet Markdown
audits.

1. TOML scope-capture (Sonnet TOML, HIGH)
   _manifests/critic-anti-pattern.toml + critic-perf.toml had [references]
   appearing AFTER [[handoff]] array-of-tables. Per TOML spec, this makes
   [references] parse as a SUB-TABLE of the last [[handoff]] element, not as
   a top-level table. All references in those manifests were silently
   unreachable by the assembler's top-level resolver.

   Moved [references] block before [[handoff]] in both files. Added 3-line
   warning comment immediately above [[handoff]] explaining the TOML scope rule
   to future editors.

2. Dangling physics-deriver in role bodies (Opus TOML, HIGH)
   Group F earlier (commit baf5425) removed [[handoff]] blocks targeting
   physics-deriver / patent-compliance / patent-researcher, but role text
   strings + forbidden_domain arrays still referenced physics-deriver in:
   - _manifests/ml-researcher.toml (lines 16, 41, 76, 89)
   - _manifests/ml-implementer.toml (line 15)
   - _manifests/infra-implementer.toml (line 16) — already scrubbed in P0
     commit 94f31ee as part of EC2-ID strip; leaving for context

   Replaced live mentions with "architect" (canonical fallback). Historical
   comments documenting the prior removal kept intentionally — they are
   documentation, not live references.

3. Wrong rule paths (Opus TOML, MEDIUM)
   ml-researcher.toml + ml-implementer.toml referenced files that don't exist
   under their stated paths:
   - path:user-rules/specialized-node-training.md → cfc-specialized-nodes.md
   - path:user-rules/observable-classification.md → paradigm-native-measurement.md

   Fixed both paths in both files.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

_manifests/critic-anti-pattern.toml

@@ -21,15 +21,18 @@ domain_in = ["task scope (verbatim user prompt)", "target paths / files"]
 forbidden_domain = ["hardcoded secrets (RULE 0.8)", "cross-language drift (use the matching sibling)"]
 output_extra_fields = ["Largest file LOC", "Tests pass count"]
 
-[[handoff]]
-target = "validator"
-trigger = "general fact-check fallback"
 [references]
 extra = [
     "path:user-rules/code-style.md",
     "path:user-rules/karpathy-behavioral.md",
 ]
 
+# IMPORTANT (TOML spec): all [table] sections must precede [[array]] sections.
+# Adding a new [section] BELOW [[handoff]] makes it a sub-table of the last handoff.
+[[handoff]]
+target = "validator"
+trigger = "general fact-check fallback"
+
 [taxonomy]
 kingdom = "manifest"
 mechanism = "compose"

_manifests/critic-perf.toml

@@ -21,15 +21,18 @@ domain_in = ["task scope (verbatim user prompt)", "target paths / files"]
 forbidden_domain = ["hardcoded secrets (RULE 0.8)", "cross-language drift (use the matching sibling)"]
 output_extra_fields = ["Largest file LOC", "Tests pass count"]
 
-[[handoff]]
-target = "validator"
-trigger = "general fact-check fallback"
 [references]
 extra = [
     "path:user-rules/code-style.md",
     "path:user-rules/karpathy-behavioral.md",
 ]
 
+# IMPORTANT (TOML spec): all [table] sections must precede [[array]] sections.
+# Adding a new [section] BELOW [[handoff]] makes it a sub-table of the last handoff.
+[[handoff]]
+target = "validator"
+trigger = "general fact-check fallback"
+
 [taxonomy]
 kingdom = "manifest"
 mechanism = "compose"

_manifests/ml-implementer.toml

@@ -12,7 +12,7 @@ role = """
 You are a senior ML implementation engineer. You write training scripts, inference code, Modal jobs, \
 and experiment runners, enforcing Math-First (Level 0), the Pre-Experiment Check, and the \
 Modal Protocol on every paid run. You own experiment observability and immediate result logging. \
-You are NOT a theory writer (hand off to `physics-deriver`), NOT a generic code writer (hand off to \
+You are NOT a theory writer (hand off to `architect`), NOT a generic code writer (hand off to \
 `code-implementer`), NOT a deploy/infra engineer (hand off to `infra-implementer`). Your output is \
 tested training/inference code with exact param counts, displayed cost estimates, and results already \
 logged in `memory/{project}.md` before analysis.
@@ -107,9 +107,9 @@ trigger = "multi-node composition design, experiment matrix layout, benchmark/ba
 [references]
 extra = [
     "path:user-rules/ml-protocol.md",
-    "path:user-rules/specialized-node-training.md",
+    "path:user-rules/cfc-specialized-nodes.md",
     "path:user-rules/api-cost-guard.md",
-    "path:user-rules/observable-classification.md",
+    "path:user-rules/paradigm-native-measurement.md",
     "path:user-rules/manifold-tangent-sanity.md",
     "path:user-rules/no-downgrade-constructive.md",
     "path:user-memory/wrong-paths-specialized-ml.md",  # TODO verify path:user-memory exists in assembler resolver

_manifests/ml-researcher.toml

@@ -13,7 +13,7 @@ You are the ML/physics research specialist. You own literature review, tooling-r
 search, reproducibility audit, and math-first formulation for any ML/RL \
 question. You are READ-ONLY — you never run experiments, never train models, never \
 edit code. Reuse beats reinvention; math beats vibes; synthetic-to-real gap is always \
-disclosed. You hand off to `ml-implementer` for experiments, `physics-deriver` for \
+disclosed. You hand off to `ml-implementer` for experiments, `architect` for \
 theorem writing, `validator` for citation gating.
 """
 
@@ -38,7 +38,7 @@ domain_in = [
 
 forbidden_domain = [
     "Running experiments, training models, or editing code (read-only agent — hand off to `ml-implementer`)",
-    "Writing theorems / derivations (hand off to `physics-deriver`)",
+    "Writing theorems / derivations (hand off to `architect`)",
     "Recommending code BEFORE writing the math expression (Math-First violation)",
     "Proposing a custom env / training loop / dataset loader without first searching MyoSuite / Menagerie / CleanRL / HuggingFace / Ninapro",
     "Reporting a sim/benchmark number without the synthetic-to-real disclaimer",
@@ -91,8 +91,8 @@ trigger = "question is about ML-system architecture (node graph, data-flow, modu
 [references]
 extra = [
     "path:user-rules/ml-protocol.md",
-    "path:user-rules/specialized-node-training.md",
+    "path:user-rules/cfc-specialized-nodes.md",
-    "path:user-rules/observable-classification.md",
+    "path:user-rules/paradigm-native-measurement.md",
     "path:user-rules/api-cost-guard.md",
     "path:user-rules/no-downgrade-constructive.md",
     "path:user-memory/wrong-paths-specialized-ml.md",  # TODO verify path:user-memory exists in assembler resolver