feat: import adversarial topology benchmark

docs/benchmarks/graph_scaling/ADVERSARIAL_TOPOLOGY_GENERALIZATION.md

@@ -0,0 +1,65 @@
+# Adversarial Topology Generalization
+
+This is the v0.5 experimental task for TS-Benchmarks.
+
+## Framing
+
+v0.4 passed on known graph families. v0.5 tries to falsify that result by testing mixed, near-boundary, and contradiction-placement-stressed graphs. The goal is not to prove universal scaling; the goal is to measure where topology-aware policy selection generalizes and where it fails.
+
+## Graph Families
+
+- `mixed_core_periphery`: scale-free hub core, small-world/random periphery, sparse bridge edges.
+- `hub_threshold_sweep`: gradually varied hub concentration near selector boundary.
+- `topology_noise_sweep`: known topology perturbed by edge rewiring/noise.
+
+## Contradiction Placements
+
+- `hub_hub`
+- `hub_leaf`
+- `leaf_leaf`
+- `random`
+
+## Oracle Comparison
+
+- `selector_policy`: chosen from pre-run topology diagnostics only.
+- `oracle_best_policy`: chosen after seeing policy outcomes.
+
+The receipt reports both, but the selector must never use oracle information.
+
+## Minimum Metrics
+
+- `selector_policy`
+- `selector_reason`
+- `oracle_best_policy`
+- `selector_matches_oracle`
+- `selector_regret_final_tension`
+- `selector_regret_f1`
+- `catastrophic_regression`
+- `boundary_case`
+- `contradiction_placement`
+- `topology_family`
+
+## Command
+
+```bash
+python -m thinking_system.benchmarks.graph_scaling.runners.adversarial_topology_generalization \
+  --sizes 100,1000 \
+  --families mixed_core_periphery,hub_threshold_sweep,topology_noise_sweep \
+  --placements hub_hub,hub_leaf,leaf_leaf,random \
+  --hub-strengths 0.25,0.35,0.45,0.60,0.75 \
+  --seed 42 \
+  --out-dir artifacts/adversarial-topology
+```
+
+## Strict Success Criteria
+
+- selector beats or matches always-reference on aggregate
+- selector beats or matches always-degree-normalized on aggregate
+- no catastrophic family-level regression
+- receipts include oracle-best-policy comparison
+- selector uses topology diagnostics only, not oracle labels or final outcome metrics
+- boundary cases are documented honestly
+
+## Claim Boundary
+
+> v0.5 is an adversarial synthetic selector stress test. It does not prove TS-Core scales to real knowledge/provenance graphs.

docs/migration/import-ledger.md

@@ -53,7 +53,7 @@ Empty placeholder directories (no tracked source): `packages/ts-*`, top-level `e
 |------------|-------------------------------------|----------------------------------|--------|-------|
 | TS-Core | `3ef48ad00efef8659ad0981d71de509a9827f584` | `src/thinking_system/core/typed_tension/` | **PARTIALLY_IMPORTED** | Domain-neutral typed tension kernel imported with its focused test and example; legacy Python application layers, Rust acceleration, CLI, Z3, Grok, UI, and Kernel Wave 12 surfaces remain unimported |
 | cig-ts-engine | `71d91a6a0cdc9e7a5439df972130596f0bae5d2f` | `src/thinking_system/graph/cig/` | **PARTIALLY_IMPORTED** | Core graph models, deterministic propagation, edge tension, derivative observer, representational radius, context-split proposal, YAML IO, seed graph and focused tests imported; CLI, plotting, reports, Obsidian export and generated artifacts remain outside |
-| TS-Benchmarks | `7c0611f7ef4dc2d150b12dc4197a8d48a462207f` | `src/thinking_system/benchmarks/graph_scaling/` | **CONSOLIDATED** | Complete implemented Workstream A imported with graph generators, relaxation policies, baselines, receipts, schemas, runners, reports, failure decomposition, topology selection, focused tests, technical documentation and retained evidence; only repository scaffolding and a public-post draft were excluded |
+| TS-Benchmarks | `7c0611f7ef4dc2d150b12dc4197a8d48a462207f` | `src/thinking_system/benchmarks/graph_scaling/` | **CONSOLIDATED** | Complete implemented Workstream A imported with graph generators, relaxation policies, baselines, receipts, schemas, runners, reports, failure decomposition, topology selection, focused tests, technical documentation and retained evidence; only repository scaffolding and a public-post draft were excluded; supplemental v0.5 adversarial-topology stress test imported from `ea6a944f5f4170a3ae02cfa8f93ce1bd87d31392` |
 | ts-chat-language | Not verified here | `packages/ts-language` | **INSPECTED** / **PLANNED** | TSLC-like code in `core/language/tslc.py` is in-tree monorepo code; full satellite port not evidenced |
 | TensionForge | Not verified here | `engines/tension-forge` | **PLANNED** (satellite); in-tree `inference/tension_forge/` is separate | Empty `engines/tension-forge/` |
 | Ten-SON-LM | Not verified here | `engines/tension-lm` | **PLANNED** | Empty destination |

docs/migration/sources/TS-Benchmarks.md

@@ -79,3 +79,26 @@ monorepo.
 
 Archiving the satellite remains a separate maintainer action after this import
 is merged and verified by CI.
+
+## Supplemental v0.5 adversarial-topology import
+
+- Source pull request: `BoggersTheFish/TS-Benchmarks#7`
+- Source branch: `experiment/v0.5-adversarial-topology-generalization`
+- Source commit: `ea6a944f5f4170a3ae02cfa8f93ce1bd87d31392`
+- Canonical issue: `#15`
+
+Imported:
+
+- mixed core/periphery graph generation
+- hub-threshold topology sweeps
+- topology-noise sweeps
+- hub-hub, hub-leaf, leaf-leaf, and random contradiction placement
+- selector-versus-oracle comparisons
+- selector regret metrics
+- catastrophic-regression detection
+- adversarial benchmark receipts and reports
+- focused tests
+
+The recorded source result failed its strict success criteria. The import
+preserves that negative result as a falsification finding, not a capability
+claim.