java-topology/docs/blast-radius.md

# CWE-407 Fix — First, Second, Third Order Effects & Blast Radius

**Internal only. No external distribution until all patches, tests, and white paper are complete.**

---

## The unlock

Every defect in this map is the same structural error: a list used where a set/map belongs,
inside a graph traversal loop. The fix is local, mechanical, and provably correct. But because
these defects live in compilers, build tools, package managers, and language runtimes — tools
that sit at the base of the software stack — the effects propagate upward through every layer
that was ever built with them.

The complexity change is O(V²) → O(V+E) for the hot path. At V=1000 nodes (a large Java
inference graph, a TypeScript checker cycle, a Tarjan SCC over package deps), this is a
**1000× reduction in membership-check operations** for that path. In practice most graphs are
small and the speedup is 2–10×, but the effect is real, measurable, and cumulative across
every build.

---

## First order effects — the patched tools themselves

These are direct. Each patched tool gets faster and its users see it immediately.

| Tool | Defect(s) | What gets faster |
|------|-----------|-----------------|
| **javac** | javac-0001..0005 | Type inference, dependency analysis, every Java compilation |
| **TypeScript tsc** | ts-0001..0003 | Cycle detection in module resolution and symbol merging |
| **GHC** | ghc-0001..0004 | SCC decode, codegen edge queries, register allocation, type-class checking |
| **Kotlin compiler** | kotlin-0001 | Non-expansive inheritance restriction checking |
| **Scala 3** | scala3-0001 | Constraint solving in type inference (currently O(n³)) |
| **CPython peg_generator** | cpython-0001 | Grammar SCC detection (affects CPython devs building Python itself) |
| **pip / distlib** | distlib-0001 | Dependency cycle detection during `pip install` |
| **GCC** | gcc-0001 | Johnson's algorithm in gcov coverage analysis |
| **LLVM / Clang** | llvm-0001 | Link-time optimization call graph traversal |
| **rustc** | rustc-0001..0002 | Match exhaustiveness checking, specialization graph build |
| **Maven** | maven-0001..0003 | Project dependency graph edge removal and cycle reporting |
| **CMake** | cmake-0001 | Link dependency group traversal |
| **npm arborist** | npm-0002 | Peer dep placement (`npm-0001` was NOT-A-DEFECT — `_depsSeen` already a `Set`) |
| **Cargo** | cargo-0001 | `cargo tree` display (display-only, bounded) |
| **Erlang stdlib** | erlang-0001 | `digraph:get_path`, `get_cycle`, `get_short_path` (`erlang-0002` FIXABLE-UPSTREAM — requires sltab in digraph.erl) |
| **Linux headerdep** | linux-0001 | Header dependency cycle detection (build tooling) |

### Blast radius at first order
- **Low.** All patches are local, behavioral equivalence is provable, and we have unit tests
  with exact operation counts that guard against regression.
- **Risk:** A patch that changes iteration order in SCC output could break a downstream
  consumer that assumed a specific ordering. Mitigation: test SCC output order explicitly.

---

## Second order effects — ecosystems built on the patched tools

### Java / JVM ecosystem (javac patches)

Everything compiled by javac benefits from faster type inference. This includes:

- **Spring Framework / Spring Boot** — millions of annotations processed per build
- **Apache Kafka, Hadoop, Cassandra, HBase** — large codebases with heavy generics
- **Android SDK toolchain** — every Android app build
- **Gradle / Maven builds** — CI/CD time drops globally
- **Bazel Java rules** — incremental builds get faster at the inference layer

#### JVM-based blockchain infrastructure

The javac patches propagate into every blockchain project built on the JVM. These are not
marginal systems — several are production financial infrastructure handling billions of dollars
in value daily.

| Project | Language | Role | Why javac matters |
|---------|----------|------|-------------------|
| **Hyperledger Besu** | Java | Full Ethereum execution client (EVM, P2P, state) | Entire codebase compiled with javac; own source is a HIGH scan target |
| **Hedera Hashgraph** | Java | Hashgraph consensus network (HBAR) | Full Java stack; consensus algorithm has graph traversal |
| **Corda / R3** | Kotlin | Enterprise permissioned ledger (financial institutions) | Kotlin compiles via javac; also benefits from kotlin-0001 patch |
| **Tron** | Java | Smart contract platform (TVM, DPoS) | Full Java validator stack |
| **Waves** | Scala | Smart contract platform | Scala compiles via javac; also benefits from scala3-0001 |
| **NEM / Symbol** | Java/TypeScript | Enterprise blockchain | Java core benefits from javac patches |
| **Hyperledger Fabric Java SDK** | Java | Permissioned ledger used by IBM, banks | SDK compiled with javac |
| **IOTA** | Rust | DAG-based ledger | Clean (Rust ecosystem; already confirmed) |

**Besu scan priority:** Hyperledger Besu is the only full Java Ethereum execution client
(alternatives geth/Nethermind/Erigon are Go/C#/Go — see CLEAN list). Besu compiles Solidity
to EVM bytecode, maintains a P2P peer graph, manages a Merkle-Patricia trie for state, and
runs EVM execution for every transaction. Graph traversal is endemic. Besu is flagged as a
HIGH scan target; scan deferred pending current wave completion.

**Second-order blast radius:** High surface area, low risk per change. The javac patches are
already tested against the installed JDK 21. The risk is that enterprise teams using older JDK
versions get the fix at different times, creating a fragmented rollout window. For financial
infrastructure (Corda, Besu, Hedera), the rollout coordination matters — these teams have
their own release cycles and may not pick up a JDK patch quickly.

### Python ecosystem (cpython, pip/distlib)

- **pip install** — every Python developer, every Docker build, every CI/CD pipeline
- **virtualenv, pipenv, poetry** — all vendor distlib or depend on pip
- **PyPI infrastructure** — resolver runs on the server side too
- **Conda** — uses its own solver but pip-compatible layer is affected
- **Docker Python base images** — `pip install -r requirements.txt` in Dockerfile layers
  is a global hotspot; faster dep resolution → faster Docker builds → faster CI

**pip specifically:** The distlib Tarjan SCC runs during `pip install` when detecting circular
dependencies in the candidate resolution set. For projects with large transitive dep graphs
(e.g., `pip install tensorflow`, `pip install scipy`), this is a non-trivial path.

### TypeScript / JavaScript ecosystem (tsc patches)

- **React, Angular, Vue, Next.js** — type-checked with tsc on every save and CI run
- **VS Code** — ships its own tsc fork; language server does cycle detection constantly
  during editing
- **Deno** — uses TypeScript compiler internals
- **Vite, esbuild, webpack** — type checking layer
- **npm, pnpm, yarn** — arborist patch (npm-0001/0002) affects every `npm install`

**Note on VS Code:** The language server runs tsc continuously. ts-0001/0002/0003 affect
interactive editing performance (symbol resolution latency, auto-complete lag in large
codebases). Fixing these is a user-visible UX improvement, not just a build-time win.

### Erlang ecosystem (erlang-0001 patched, erlang-0002 FIXABLE-UPSTREAM)

`digraph` and `digraph_utils` are OTP stdlib — the graph library for the entire Erlang
and Elixir ecosystem. The erlang-0002 fix (`loop_vertices/1`, `is_simple/1`: O(V²) →
O(V)) propagates to every application that calls those functions on OTP upgrade.

**Systems that get faster:**
- **RabbitMQ** — uses `digraph` for exchange routing graph validation (topology cycles,
  simplicity checks on the exchange graph during reconfiguration)
- **ejabberd** — XMPP routing graph; topology validation at cluster join
- **Rebar3** — Erlang build tool; dependency graph analysis
- **Mix / Hex** — Elixir build tool; same OTP digraph calls
- **Any Erlang/Elixir application** that calls `loop_vertices/1` or `is_simple/1`

**The speed increase is real and correct. But it carries secondary risk for
timing-sensitive systems.**

#### The throttle risk — queue-based and financial systems

Erlang is the runtime of choice for telecom infrastructure, financial messaging, and
high-throughput queue systems. Some of these systems have been capacity-planned and
operationally tuned around *observed* performance characteristics of the current OTP
runtime, including graph operations that run during topology changes.

**Risk pattern:** If `loop_vertices` or `is_simple` was running slowly enough to act
as an implicit throttle during exchange graph reconfiguration (e.g., RabbitMQ
vhost topology change, ejabberd MUC room graph validation), downstream consumers
may have been sized assuming that rate. A sudden 100×–1000× speedup in that path
changes the rate at which topology changes are processed, potentially triggering
thundering-herd behavior in systems that were never expected to handle topology
changes at that speed.

**Specific systems to audit before deploying the OTP patch:**

| System | Risk | Why |
|--------|------|-----|
| **RabbitMQ** | Medium | Exchange topology validation rate increases; downstream consumers of topology-change events need capacity review |
| **Financial messaging (LMAX Disruptor-style Erlang systems)** | Medium-High | Queue scheduling and backpressure logic may be calibrated to current digraph latency |
| **Stock exchange order routing** | High if affected | Any Erlang-based order router where exchange graph validation is in the latency-critical path must be re-benchmarked |
| **ejabberd MUC** | Low | Room graph ops are infrequent and not in the message routing hot path |
| **Rebar3 / Mix** | None | Build tooling only; faster is unambiguously good |

**The general principle:** This is the "faster is dangerous when slow was load-limiting"
problem. It applies to any OTP-based system where:
1. `loop_vertices/1` or `is_simple/1` runs during a state-change event
2. That event feeds a downstream system with a fixed processing budget
3. That downstream system was sized against the current (slow) call latency

**Mitigation:** Before deploying the OTP patch in any financial or queue-based
production system:
1. Identify all call sites of `digraph_utils:loop_vertices/1` and `is_simple/1`
   in the application and all dependencies
2. Measure the current call latency under production-representative load
3. Model the downstream effect of 100× speedup at those sites
4. Adjust backpressure, rate limiting, or consumer capacity as needed
5. Stage the rollout: canary → 10% → 100% with monitoring on downstream queue depth

### Haskell ecosystem (GHC patches)

- **Pandoc** — compiled with GHC, used everywhere for document conversion
- **Cardano** — blockchain written in Haskell; smart contract compilation is affected
- **XMonad, Yi** — Haskell tooling
- **Stack, Cabal** — build tools that invoke GHC; faster GHC = faster Haskell builds
- **ghc-0002 codegen** — every function that generates LLVM IR via GHC gets the benefit

### Rust ecosystem (rustc patches)

- **Firefox** — compiled with rustc; match exhaustiveness checker (rustc-0001) runs
  on every enum
- **ripgrep, fd, bat, exa** — popular CLI tools; faster specialization builds
- **Servo** — rendering engine in Rust
- **Cargo itself** — cargo-0001 is display-only, bounded, but still a principle violation

### Browser ecosystem

Browsers are among the largest and most performance-critical C++/Rust codebases on the
planet. All three major engines are affected by patches already in this map.

**Firefox**
- Compiled with Clang; LLVM LTO enabled in release builds → llvm-0001 applies directly
- Large Rust codebase: WebRender (GPU compositor), Stylo (CSS engine), Servo components
  → rustc-0001 (match exhaustiveness) and rustc-0002 (specialization graph) both apply
- SpiderMonkey IonMonkey JS engine: **CLEAN** (confirmed — uses hash containers throughout)
- TypeScript used in Firefox DevTools and web-ext tooling → ts-0001..0003 apply

**Chrome / Chromium**
- Compiled with Clang; LLVM LTO in all release builds → llvm-0001 applies
- V8 TurboFan JIT compiler: **CLEAN** (confirmed — no CWE-407 candidates)
- TypeScript used in DevTools, Chrome Extensions API, web platform test tooling
- Large npm dependency graph for web platform tooling → npm arborist patch applies

**Safari / WebKit**
- Compiled with Clang; LLVM LTO applies → llvm-0001 applies
- JavaScriptCore (JSC): not yet scanned; lower probability than SpiderMonkey/V8 given
  Apple's engineering culture, but a candidate
- WebKit build system uses CMake → cmake-0001 applies

**The LTO point:** All three browsers ship release builds with link-time optimization
enabled. LLVM's GlobalsModRef call-graph traversal (llvm-0001) runs across the entire
binary during LTO. Firefox is ~10M LOC, Chromium is ~35M LOC. For Chromium specifically,
LTO build time is a known bottleneck — the GlobalsModRef fix is material.

**Browser JS engine scan summary:**

| Engine | Browser | Result |
|--------|---------|--------|
| V8 TurboFan | Chrome | **CLEAN** |
| SpiderMonkey IonMonkey | Firefox | **CLEAN** |
| JavaScriptCore | Safari | Not yet scanned |

### C/C++ ecosystem (GCC, LLVM, CMake patches)

This is the broadest surface area. GCC and LLVM compile essentially everything:

- **PostgreSQL** — compiled with GCC/Clang; build time improves
- **SQLite** — compiled with GCC/Clang
- **MySQL / MariaDB** — compiled with CMake + GCC/Clang; CMake patch (cmake-0001)
  directly speeds up the MySQL build's link-dependency resolution
- **Apache httpd (apache2)** — autotools + GCC; build time improves
- **nginx** — GCC; build time improves
- **Caddy** — written in Go (Go compiler already clean ✓); no effect here
- **OpenSSL, libssl** — GCC/Clang compilation benefits
- **Linux kernel** — GCC + headerdep.pl (linux-0001) patched

**LLVM LTO specifically (llvm-0001):** Link-time optimization is used by default in release
builds of Firefox, Chrome, Rust's standard library, LLVM itself, and PostgreSQL with `--enable-lto`.
The GlobalsModRef call-graph traversal runs during LTO. For large LTO builds (Firefox is
~10M LOC), this is a meaningful path.

### Second-order blast radius summary

| Ecosystem | Risk level | Primary concern |
|-----------|-----------|-----------------|
| JVM / Android | Medium | JDK rollout fragmentation across versions |
| Python / pip | Low-Medium | pip is heavily tested; distlib change is isolated |
| TypeScript / npm | Medium | VS Code ships its own tsc; needs separate coordination |
| Haskell | Low | GHC releases are infrequent, community is small |
| Rust | Low | rustc team has strong test infra |
| C/C++ / GCC / LLVM | Medium-High | Widest surface; GCC/LLVM release cycles are long |

---

## Third order effects — infrastructure and runtime systems

### Databases

**PostgreSQL**
- Build-time: benefits from GCC/CMake fix (faster to compile from source)
- Runtime: PostgreSQL's query planner has **5 confirmed CWE-407 defects** (postgresql-0001
  through -0005) in `tlist.c`, `preptlist.c`, `equivclass.c`, `analyzejoins.c`, and `list.c`.
  All are **DEFERRED** — the standard fix (replace list with hash set) is blocked by the
  absence of a generic expression hash function in PostgreSQL core. See the executive summary
  doc for full analysis. These sites require upstream PostgreSQL core collaboration before
  a patch is possible.
- Extension ecosystem: PL/Python, PL/Perl, PostGIS — all pull in the patched language runtimes

**SQLite**
- Build-time: GCC fix applies
- Runtime: SQLite's query optimizer is simpler (no join reordering); lower risk of CWE-407
  in its own source. However, SQLite is used as an embedded DB in Python (`sqlite3` module),
  Ruby, PHP, Node.js — all of which are getting faster runtimes from our patches.

**MySQL / MariaDB**
- Build-time: CMake patch (cmake-0001) directly applies — MySQL build uses CMake heavily
- Runtime: MySQL's optimizer handles join graphs; candidate for its own CWE-407 scan

**MongoDB**
- Compiled with SCons + GCC/Clang; build improves
- Aggregation pipeline planner: candidate for scan

### Web servers

**apache2 (httpd)**
- Compiled with GCC; build-time improvement
- `mod_proxy`, `mod_rewrite` rule graphs: low complexity, bounded inputs
- Apache Traffic Server (ATS) — more complex routing graph; candidate for scan

**nginx**
- Compiled with GCC; build-time improvement
- nginx config parsing is linear, low-complexity graph work; low risk of CWE-407 in nginx itself

**Caddy**
- Written in Go — Go compiler already clean (✓ in our clean list)
- Caddy's own routing graph uses Go maps throughout; low risk

### GeoIP and geographic routing — special considerations

This is the most subtle third-order effect.

**The issue:** GeoIP accuracy is imperfect. GeoIP databases (MaxMind GeoLite2, IP2Location,
etc.) have known error rates — typically 95-99% accurate at country level, 60-80% at city
level. These errors cause:
- Misrouted CDN requests (user in Frankfurt hits London PoP)
- Payment fraud false positives (billing address country ≠ detected country)
- Content geo-restrictions misfiring

**The connection to our fix:** Geographic load balancers and CDN routing systems are built on
top of the language runtimes we're patching. Faster compilation and package resolution means:
1. Routing rule updates deploy faster → errors propagate faster too
2. If a GeoIP correction patch ships faster (because pip install / npm install / maven build is
   faster), it reaches production faster — which is good when the correction is right, and
   propagates faster when the correction itself contains an error

**Specific risk:** MaxMind's geoip2 Python library runs on CPython. If CPython's own build
tooling improves (cpython-0001 is in peg_generator, used when regenerating the parser), and
pip's dep resolution improves (distlib-0001), then GeoIP library updates reach production
faster. At scale (millions of IPs routed per second), even a brief incorrect GeoIP DB update
is amplified.

**Mitigation:** GeoIP database deployments should be blue/green with traffic validation at 1%
before full rollout — this is good practice regardless of our patches but becomes more
important as deployment velocity increases.

**The broader geo paradox:** Our fix makes the whole stack faster. Faster stacks reduce
latency. Reduced latency can shift requests between geographic regions (requests that were
timing out now succeed, from further away). This very slightly shifts the apparent distribution
of traffic origins — which feeds back into GeoIP accuracy metrics. This is a genuine second-
order effect that geo-aware systems (ad targeting, fraud detection, CDN) should be aware of.

### CI/CD and cloud infrastructure

**Docker image builds**
- Python base images: `pip install` in Dockerfile layers is the single biggest time sink
  in most CI pipelines. distlib-0001 + cpython-0001 together reduce this.
- Maven/Gradle builds: Java CI pipelines benefit from javac patches
- npm install: arborist patches reduce dep tree construction time

**At scale:** GitHub Actions processes ~50M workflow runs/month. If each Java/Python/TS
workflow saves 5-15 seconds of build time, the aggregate is millions of compute-hours/month.
This is real money and real carbon.

**Serverless cold starts**
- AWS Lambda Python runtime: pip-installed dependencies ship in the Lambda layer;
  faster resolution = smaller/simpler layers = faster cold starts
- Lambda Java runtime: javac inference improvements are baked into compiled JARs — no direct
  cold-start benefit, but the JIT (C2 compiler, which uses LLVM-like graph analysis) may benefit
  indirectly

---

## Blast radius mitigation plan

### Tier 1 — Before any patch is submitted upstream

1. **Every patch has a behavioral equivalence proof** — not just "tests pass" but a written
   argument that output is identical for all inputs (SCC membership, ordering, cycle reporting)
2. **Operation-count unit tests** — if a test does not assert O(1) membership, it does not count
3. **Fuzz testing on graph structure** — random DAGs, random dense graphs, self-loops,
   disconnected components, very large graphs (V=10,000+)
4. **No patch touches error messages or exception types** — changing "contains" to a set
   must not change what gets thrown or printed when a cycle is detected

### Tier 2 — Before coordinated disclosure

5. **Upstream maintainer contact before public patch** — privately share the patch and proof
   with the maintainer; give them 90 days to merge and release
6. **Sequence disclosure by blast radius** — patch low-surface tools first (peg_generator,
   distlib, erlang stdlib) before high-surface tools (javac, tsc, GHC)
7. **Version compatibility testing** — test each patch against the last 3 major releases of
   the affected tool, not just HEAD

### Tier 3 — Infrastructure-specific

8. **Database query planners** — scan PostgreSQL, MySQL, MongoDB source for CWE-407 before
   disclosing compiler fixes. If db planners have the same defect, coordinate disclosure
   together. Revealing "compilers are fixed" while "your DB planner has the same bug" creates
   an exploit window.
9. **GeoIP deployment velocity** — note in the white paper that faster deployment pipelines
   increase the importance of staged rollouts for data updates (not just code)
10. **CDN / routing system operators** — brief major CDN operators (Cloudflare, Fastly, Akamai)
    as part of coordinated disclosure. Their build pipelines are affected; their traffic routing
    systems may independently contain the same defect.

### Tier 4 — Post-disclosure monitoring

11. **Regression watch** — monitor upstream repos for 6 months post-disclosure for any
    performance regression reports that could be attributed to ordering changes in SCC output
12. **CVE coordination** — CWE-407 in a build tool is typically a DoS via crafted input
    (an adversary can construct a source file that maximizes the quadratic behavior). File CVEs
    for tools that accept untrusted input (tsc, javac, GCC/Clang — all accept user source).
    Do NOT file CVEs for internal-only tools where input is trusted.

---

## Fourth frontier: scientific computing and numerical analysis

This is the domain where the topology defect may be causing the most *invisible* damage.
Scientific computing works on genuinely large graphs — protein interaction networks (V=20,000+),
genomics dependency graphs, finite element meshes, neural computation graphs, Monte Carlo
dependency chains. At these scales, O(V²) is not "a bit slow" — it is **computationally
unobservable**. Researchers simply never run the algorithm on the full dataset; they subsample,
they approximate, they accept that "large graphs are slow."

### NetworkX (Python)

NetworkX is the dominant pure-Python graph library. Used by: bioinformatics, social network
analysis, quantum circuit simulation, ML pipeline graphs, physics simulations.

NetworkX implements Tarjan SCC, Kosaraju SCC, DFS, topological sort, cycle detection,
dominator trees, and dozens of other graph algorithms entirely in Python. The entire library
predates the widespread adoption of O(1)-first idioms in Python graph code.

**High-probability CWE-407 targets in NetworkX:**
- `networkx/algorithms/components/strongly_connected.py` — Tarjan/Kosaraju SCC
- `networkx/algorithms/cycles.py` — `simple_cycles()`, `find_cycle()`
- `networkx/algorithms/dag.py` — topological sort, cycle detection
- `networkx/algorithms/dominance.py` — dominator tree construction
- `networkx/algorithms/traversal/depth_first_search.py` — DFS with visited tracking

If any of these use `list` for the visited/stack/path set, every scientific computing
workflow that calls them on large graphs has been running at O(V²) instead of O(V+E).
At V=10,000, that is a 10,000× error in expected runtime. Researchers would observe this
as "NetworkX doesn't scale" and switch to igraph or switch languages — never knowing the
fix was one data structure change.

**Scan priority: CRITICAL.** NetworkX is installed in virtually every scientific Python
environment. Clone and scan immediately.

### SciPy `csgraph`

`scipy.sparse.csgraph` implements Dijkstra, Bellman-Ford, Floyd-Warshall, minimum spanning
tree, connected components, and shortest paths. The core algorithms are written in Cython
and compiled to C — the hot paths are likely clean. But:

- The Python dispatch layer wraps these C routines and may do list-based bookkeeping
- `scipy.sparse.csgraph.depth_first_order` — DFS with predecessor/successor arrays;
  the Python-level visited set tracking is a candidate
- `scipy.sparse.csgraph.minimum_spanning_tree` — Kruskal uses union-find (clean) but
  Prim variants may not

SciPy is used in: finite element analysis, fluid dynamics simulation, computational
chemistry, signal processing pipelines. Wrong graph complexity at this layer means
numerical simulations are taking longer than the physics requires.

### NumPy

NumPy itself does not implement graph algorithms. But NumPy arrays are frequently used as
adjacency matrices fed into graph libraries — and the conversion layer (numpy array →
graph structure) may introduce O(n) membership patterns. Lower priority than NetworkX,
but the `numpy.lib.arraysetops` and `numpy.unique` functions used in graph preprocessing
pipelines are worth scanning for misuse.

### igraph (C core + Python/R bindings)

igraph is a C library widely used in network science and bioinformatics. The C core is
likely clean (C programmers tend to use arrays with boolean flags). But the Python and R
binding layers, and the igraph R package's pure-R wrappers, are candidates.

### Graph-ML frameworks

- **PyTorch Geometric (PyG)** — graph neural networks. Uses Python-level graph traversal
  for neighborhood sampling, subgraph extraction, message passing setup.
- **DGL (Deep Graph Library)** — similar. Graph partitioning and traversal in Python layer.
- **TensorFlow graph executor** — C++. Execution graph SCC and topological sort are
  internal; likely clean (Google engineers), but worth scanning.
- **JAX** — computation graph tracing in Python. `jax.core` builds and traverses Jaxpr
  graphs during tracing. Candidate for CWE-407 in the trace/compilation path.

**The ML training implication:** If graph traversal in a GNN framework's data loading or
batching code is O(V²), large-graph training runs that appear to stall at the data
preparation stage may actually be fixable with a one-line patch. This would directly
reduce training costs at scale.

### Numerical analysis: the "wrong answer" risk

This is the most serious concern beyond performance. Some numerical algorithms use graph
traversal to determine computation order (e.g., sparse matrix factorization, automatic
differentiation, constraint propagation). If the traversal produces a **different ordering**
due to a latent defect — not a performance defect but an **ordering defect** — the
numerical results themselves could be subtly wrong.

Example: sparse Cholesky factorization uses a fill-reduction ordering step (AMD, METIS)
that involves graph traversal. If a visited-set defect causes a node to be processed twice
or skipped, the fill pattern changes. The factorization still "works" but has higher fill
than optimal, consuming more memory and producing different round-off error patterns.

This is speculative but must be ruled out. The white paper needs a section specifically
addressing whether any of the defects in our map could produce incorrect output (not just
slow output) under any input. Current assessment: **no** — all confirmed defects degrade
to O(n²) but produce correct output. But numerical computing chains using these libraries
must be individually verified.

---

## Fifth frontier: network routing protocols and infrastructure

This is where the topology defect ceases to be a software quality issue and becomes a
**live infrastructure reliability issue.**

Network routing protocols are graph algorithms running continuously on production hardware,
reacting to topology changes in real time. If their graph traversal has quadratic membership
checks, the convergence behavior of the internet itself is degraded relative to theoretical
bounds.

### BGP (Border Gateway Protocol)

BGP is the routing protocol of the internet — it maintains reachability between all
autonomous systems (ASes). BGP routers maintain route tables with 900,000+ IPv4 prefixes
and process updates continuously.

**AS-path loop detection:** BGP prevents routing loops by checking if the local AS number
appears in the AS-path of an incoming route. In a naive implementation, this is a linear
scan of the path list. For typical paths (4-8 ASes) this is negligible. But:

- During BGP route storms (mass withdrawal + re-advertisement), a router may process
  millions of updates/second
- If the loop detection iterates a list rather than checking a set/bitmap, the cost per
  update multiplies with path length
- Route reflectors in large ISP networks see paths of 20-50 ASes for international routes

**FRRouting (FRR)** — the most widely deployed open-source BGP/OSPF/IS-IS implementation.
Used by major cloud providers (Meta, Microsoft, LinkedIn use FRR derivatives). Written in C.
`bgpd/bgp_aspath.c` — AS-path manipulation and loop detection. **High-priority scan target.**

**BIRD** — widely used in IXP (Internet Exchange Point) route servers. Written in C.
`proto/bgp/` — BGP implementation. Candidate.

**ExaBGP** — Python BGP implementation used for route injection and traffic engineering.
Pure Python. If its path-traversal code uses list membership, it inherits the defect
directly. **Very high probability of CWE-407 given the language.**

**GoBGP** — Go implementation. Go uses maps natively. Likely clean. Worth verifying.

**OpenBGPD** — OpenBSD BGP daemon. C. BSD codebase tends to be careful but written in
pre-modern-idiom era. Candidate.

### OSPF (Open Shortest Path First)

OSPF runs Dijkstra's Shortest Path First (SPF) algorithm on the link-state database. SPF
is triggered every time the topology changes (link up/down, metric change). On large
networks (enterprise core, ISP backbone), SPF runs on graphs of hundreds to thousands
of nodes.

**If Dijkstra's visited set is a list:** O(V²) per SPF run. OSPF specs require SPF to
complete in milliseconds. A quadratic implementation on a 1000-node network doing 1M
operations instead of 1000 would fail to meet convergence timers, causing route
oscillation and potentially forming routing black holes during convergence.

**FRRouting `ospfd`** — `ospfd/ospf_spf.c`. SPF implementation in C. **Critical scan target.**
If this has CWE-407, it is a live network reliability defect in production ISP infrastructure.

**The convergence timer implication:** OSPF defines `SPF_DELAY` (default 200ms) and
`SPF_HOLDTIME` (default 1000ms). If SPF takes longer than expected due to quadratic
behavior, the hold-time backs off and convergence slows — making the network appear
to be "under load" when it is actually hitting a complexity defect.

### IS-IS (Intermediate System to Intermediate System)

IS-IS is the other major link-state IGP, preferred by many large ISPs (Google, Comcast)
and most carrier backbone networks. Also uses SPF. Same risk as OSPF.

**FRRouting `isisd`** — `isisd/isis_spf.c`. Candidate.
**CLNS IS-IS in IOS/IOS-XR** — Cisco's implementation, closed-source, cannot scan directly.
But if FRR has the defect, Cisco almost certainly inherited similar code from the same
1990s-era algorithm literature.

### MPLS and traffic engineering

MPLS label-switched paths are computed using RSVP-TE or SR-TE path computation. Path
computation involves constrained shortest-path first (CSPF) — Dijkstra with constraints.
CSPF runs on a graph of the entire network for each LSP setup. If visited-set is a list:
O(V²) per tunnel setup. In a network with thousands of MPLS tunnels being re-signaled
after a failure, this would cause a tunnel re-establishment storm.

**OpenDaylight (ODL)** — Java SDN controller. Implements PCE (Path Computation Element)
for MPLS-TE. Java + graph algorithms = **very high probability of CWE-407**. Used by
major telcos for network automation. Scan target.

**ONOS (Open Network Operating System)** — Java SDN controller used by AT&T, NTT, Comcast.
`core/api/src/main/java/org/onosproject/net/topology/` — topology service. Java list-based
graph traversal almost certain. **High-priority scan target.**

### Network middleware and service meshes

**HAProxy** — C, load balancer. Route selection is simple weighted round-robin or least-conn,
not graph-based. Low risk of CWE-407 in its own code. But HAProxy's configuration validator
may parse ACL dependency graphs — candidate for the config-parse path.

**Envoy Proxy** — C++, service mesh. Envoy's cluster graph, endpoint discovery, and routing
rule evaluation are all graph-structured. The xDS API builds a runtime graph of clusters,
endpoints, and listeners. `source/common/upstream/` — cluster dependency resolution.
**Medium-priority scan target.**

**Istio (control plane)** — Go. Pilot builds an Envoy configuration graph and pushes it.
Go-based, likely uses maps. But `pilot/pkg/networking/core/` — virtual service graph
resolution. Worth verifying.

**Consul** — Go service mesh. Likely clean.

**Linkerd** — Rust. Very likely clean.

**Cilium** — Go + eBPF. Policy graph in Go. Likely clean.

### The internet reliability implication

If FRR's OSPF SPF or BGP path-selection has a quadratic membership check:

1. **Every network failure event** triggers slower-than-specified convergence
2. **BGP route storms** (which happen regularly at major IXPs) cause CPU spikes that are
   currently attributed to "BGP flapping load" but may actually be algorithmic overhead
3. **The internet's recovery time from fiber cuts, hardware failures, and DDoS attacks
   is longer than it needs to be** — not by a little, but potentially by orders of
   magnitude on large networks

This is not hypothetical. There are documented cases of BGP convergence taking minutes
instead of seconds on large networks. The standard explanation is "BGP is slow by design."
The actual explanation may include quadratic graph traversal.

### Routing scan backlog — immediate priority

| System | Language | Key file to scan | Why critical |
|--------|----------|-----------------|--------------|
| FRRouting bgpd | C | `bgpd/bgp_aspath.c`, `bgpd/bgp_route.c` | AS-path loop detection |
| FRRouting ospfd | C | `ospfd/ospf_spf.c` | Dijkstra SPF on every topology change |
| FRRouting isisd | C | `isisd/isis_spf.c` | IS-IS SPF, carrier backbone |
| FRRouting ldpd | C | `ldpd/lde_lib.c` | MPLS label distribution |
| BIRD bgp | C | `proto/bgp/bgp.c` | IXP route servers globally |
| ExaBGP | Python | `exabgp/bgp/message/update/attribute/aspath.py` | Pure Python, high probability |
| OpenDaylight PCE | Java | `pcep/` topology service | MPLS-TE path computation |
| ONOS topology | Java | `core/net/src/main/java/org/onosproject/net/topology/` | Major telco production |
| NetworkX | Python | `algorithms/components/`, `algorithms/cycles.py` | Scientific computing globally |
| ExaBGP | Python | all graph traversal paths | Very high CWE-407 probability |

---

## Sixth frontier: MATLAB, CAD, and engineering simulation software

### MATLAB

MATLAB is the primary computational tool for control systems, signal processing, circuit
simulation, and numerical methods in engineering. MathWorks ships MATLAB with a graph
and network algorithms toolbox and the core language runtime includes graph traversal
in multiple places.

**MATLAB's own graph algorithms (`graph` / `digraph` objects, introduced R2015b):**
- `conncomp()` — connected components (SCC for directed graphs)
- `toposort()` — topological sort with cycle detection
- `shortestpath()`, `shortestpathtree()` — Dijkstra/Bellman-Ford
- `isdag()` — cycle detection

These are implemented in MATLAB's compiled C/C++ runtime (MathWorks closed source).
Cannot scan directly. But the behavioral signatures are observable: benchmark
`conncomp(G)` on random digraphs as V grows. If runtime is O(V²) instead of O(V+E),
the defect is present.

**MATLAB's build/dependency tooling:**
- Simulink uses a signal-flow graph to determine block execution order. Block sorting is
  topological sort. If the visited set in that sort uses MATLAB cell array membership
  (the MATLAB equivalent of list membership — O(n) via `ismember()`), every Simulink
  model compilation has this defect.
- MATLAB's `depfun()` and the newer `matlab.codetools.requiredFilesAndProducts()` build
  dependency graphs. If linear membership is used, large codebases are slower than necessary.

**MATLAB m-file graph code:** The MATLAB community writes enormous amounts of graph
algorithm code in `.m` files. `ismember(x, list)` in MATLAB is O(n) by default (it sorts
and binary-searches, so O(n log n) for the sort + O(log n) query — better than naive O(n)
but still not O(1)). Code that uses `ismember()` in a DFS loop is O(V·n·log n). The
idiomatic fix is `containers.Map` (hash map) or logical indexing arrays.

The MATLAB file exchange (100,000+ submissions) and most academic graph theory `.m` files
predate idiomatic O(1) membership in MATLAB. The defect is endemic in the research codebase.

**Octave** (open-source MATLAB-compatible): same patterns, scannable. GNU Octave was scanned
(2026-03-23) and is **CLEAN** — all graph algorithms use vectorized ops and compiled C routines,
not list-backed visited sets. The MATLAB `ismember` risk applies to user-authored `.m` files,
not Octave's own implementations.

### Simulink and Model-Based Design

Simulink is used to design control systems for aircraft, automobiles, medical devices, and
industrial machinery. The compiled output (via Embedded Coder) runs in safety-critical
hardware. The design-time graph traversal (execution order, algebraic loop detection,
rate transition analysis) uses the MATLAB runtime.

**Algebraic loop detection** is Tarjan SCC on the block diagram graph. If this runs at
O(V²), large Simulink models (aerospace, automotive — common at V=10,000 blocks) are
taking far longer to compile than necessary. Engineers accept slow model compilation as a
fact of life. It may not be.

**DO-178C / ISO 26262 implication:** If the model compiler has a performance defect, it
affects the model development cycle time but not (directly) the correctness of generated
code. However, if Simulink's cycle detection produces incorrect results due to an ordering
defect (not just a performance defect), that is a safety-critical issue. This must be
ruled out explicitly.

### CAD / EDA (Electronic Design Automation)

EDA tools are the compilers of hardware. They process netlists — graphs of logic gates,
wires, and timing constraints — and produce manufacturable chip designs. The graph
algorithms in EDA are among the most performance-critical in all of engineering.

**Key graph algorithms in EDA:**
- **Technology mapping** — covering a DAG of logic operations with library cells (graph
  covering, DFS-based)
- **Static timing analysis (STA)** — longest path in a DAG (topological sort + DP)
- **Place and route** — graph partitioning, Steiner tree, maze routing
- **Equivalence checking** — SCC-based circuit comparison
- **Power analysis** — reachability in switching activity graph

**Open-source EDA tools (scannable):**

| Tool | Language | Graph algorithm | Scan target |
|------|----------|----------------|-------------|
| **Yosys** | C++ | synthesis, technology mapping, SCC | `passes/opt/`, `kernel/rtlil.cc` |
| **OpenROAD** | C++ | placement, routing, timing | `src/odb/`, `src/sta/` |
| **OpenSTA** | C++ | static timing analysis, DAG traversal | `graph/`, `search/` |
| **ABC** (Berkeley) | C | logic synthesis, DAG rewriting | `src/base/abci/` |
| **VPR** (Verilog-to-Routing) | C++ | FPGA place and route | `vpr/src/route/` |
| **Icarus Verilog** | C++ | netlist elaboration, dependency graph | `tgt-vvp/` |
| **Verilator** | C++ | RTL simulation, SCC for clock domains | `src/V3Graph.cpp` |

**Verilator specifically:** `V3Graph.cpp` and `V3GraphAlg.cpp` are Verilator's internal graph
library. Used for SCC computation on hardware description language graphs. Verilator is used
by Google, lowRISC, and chip startups to verify RISC-V designs. If its SCC has a quadratic
membership check, large chip verification runs are slower than necessary.

**Commercial EDA (Cadence, Synopsys, Mentor):** Closed source, cannot scan. But the same
algorithm literature was used. Performance benchmarks of commercial tools on large netlists
may reveal the signature of quadratic behavior.

**The chip design implication:** EDA tool runtime directly determines chip design cycle time.
Longer compile times → fewer design iterations → worse final chip quality. If O(V²) graph
traversal is embedded in EDA tools, the chips being designed today are suboptimal relative
to what the tools could produce with correct complexity.

### Other CAD / simulation systems

**FreeCAD / OpenCASCADE** — C++. Parametric dependency graph for feature ordering. If
feature rebuild order uses list-based visited tracking, complex assemblies rebuild slowly.

**KiCad** — C++. PCB netlist graph for DRC (design rule check) and copper pour. DRC
involves connectivity analysis. `pcbnew/connectivity/` — scan target.

**Blender** — C/Python. Node graph compositor and geometry nodes use topological sort for
execution order. `source/blender/blenkernel/intern/node.cc` — dependency graph. Candidate.

**FEniCS / OpenFOAM** — finite element / computational fluid dynamics. Build mesh adjacency
graphs. Python and C++ layers. Mesh partitioning algorithms involve graph traversal.

---

## Backlog — systems not yet scanned, priority order

**Confirmed CLEAN (scanned 2026-03-23, no action needed):**
ONOS, OpenDaylight, MySQL optimizer, V8 TurboFan, SpiderMonkey IonMonkey, Bazel, sbt,
GNU Octave, KiCad, Yosys, Verilator — all use O(1) hash containers for graph traversal state.

**PostgreSQL (5 sites DEFERRED):** Scanned and confirmed defective; patch blocked by missing
`nodeHash()` infrastructure. Requires upstream collaboration. See executive summary.

**Remaining unscanned — priority order:**

| System | Language | Why critical | Scan approach |
|--------|----------|-------------|---------------|
| **FRRouting ospfd SPF** | C | `ospf_spf.c` Dijkstra — live router convergence (TI-LFA already patched) | scan (C) |
| **FRRouting bgpd** | C | `bgp_aspath.c` AS-path loop detection | scan (C) |
| **FRRouting isisd** | C | `isis_spf.c` IS-IS SPF, carrier backbone | scan (C) |
| **NetworkX** | Python | Scientific graph library; academic/research global baseline | clone + scan (Python) |
| **ExaBGP** | Python | Pure Python BGP; very high probability | clone + scan (Python) |
| **OpenSTA** | C++ | Static timing analysis | clone + scan (C++) |
| **Blender node graph** | C/Python | Geometry nodes, compositor | clone + scan |
| **BIRD bgp** | C | IXP route servers | clone + scan (C) |
| **OpenBGPD** | C | BSD BGP | clone + scan (C) |
| **Buck2** | Rust | Build target graph | clone + scan (Rust) |
| **Pants** | Python | Build target graph | clone + scan (Python) |
| **NuGet** | C# | .NET dep resolution | clone + scan (C#) |

---

## Summary

The paradigm unlock is real and the blast radius is broad, but it is manageable because:

1. The fix is **local** — one data structure change per site, no algorithm redesign
2. The fix is **provably correct** — set/map membership is semantically identical to list membership for these use cases; only the complexity changes
3. The fix is **testable** — operation counts are measurable and we can assert them in tests
4. The blast radius grows upward through layers that are **already heavily tested** — we are not inserting new behavior into PostgreSQL or nginx; we are making their build tools faster

**The routing/scientific computing concern is qualitatively different from the compiler concern.**
For compilers and build tools, the defect causes slow builds. For network routing protocols,
the defect may be causing live convergence failures. For scientific computing, the defect may
be causing researchers to accept wrong performance baselines and design experiments around
them. For EDA tools, it may be extending chip design cycles. These are not "software quality"
issues — they are infrastructure reliability and scientific integrity issues.

The geo/CDN/routing concern is real but indirect: our fix accelerates deployment pipelines,
which amplifies both good updates and bad ones. The mitigation is staged rollout discipline
in those systems — which is good practice regardless of our work.

**Scope of work is larger than initially mapped.** The 35-site defect map covers compilers
and build tools. Routing protocols, scientific computing, EDA, and numerical simulation are
a separate wave — same defect pattern, different domain, potentially higher real-world impact.

---

## Seventh frontier: web infrastructure stack

This layer sits between the internet and application code. Every HTTP request passes through
one or more of these systems. Graph algorithms appear in: module dependency resolution,
request routing rule evaluation, VCL/config compilation, PHP opcode compilation, and
upstream cluster topology management.

### PHP — Zend Engine (HIGH PROBABILITY)

PHP's Zend Engine compiles PHP source to opcodes at runtime. The compilation pipeline is a
classic compiler pipeline with full graph algorithm infrastructure:

- **`Zend/zend_cfg.c`** — Control Flow Graph (CFG) construction. Basic block discovery,
  predecessor/successor lists. DFS-based.
- **`Zend/zend_dfg.c`** — Data Flow Graph. Liveness analysis, reaching definitions.
  Iterative dataflow over CFG — the fixed-point loop visits nodes and may check
  visited/changed state with list membership.
- **`Zend/zend_ssa.c`** — SSA (Static Single Assignment) form construction. Requires
  dominator tree, dominance frontiers — classic graph algorithms.
- **`Zend/zend_optimizer.c`** — Optimizer over SSA form. DCE, SCCP, type inference.

PHP is executed on every web request (unless opcode cached). The opcode cache
(OPcache) means CFG/SSA is built once per file, not per-request. But any CWE-407 in
CFG/SSA construction affects every PHP deploy's warm-up time and memory usage.

**`ext/opcache/Optimizer/`** — OPcache optimizer passes. Multiple graph traversal passes.
`zend_ssa.c`, `zend_call_graph.c`, `zend_func_info.c` are all graph algorithm files.
**Highest-probability CWE-407 target in this tier.**

**PHP Composer** — the PHP package manager. Pure PHP dependency resolver.
`src/Composer/DependencyResolver/` uses a pool-based SAT solver with graph operations.
If any visited/cycle set is a PHP array (O(n) `in_array()`), it inherits distlib-0001's
pattern in the PHP ecosystem.

### Redis

Redis uses graph structures in:
- **Cluster topology** (`src/cluster.c`) — cluster nodes maintain predecessor/follower
  state, path-finding for slot migration, and reachability for failover detection.
  Redis Cluster's `clusterGetSlotByQuery`, `clusterNodeGetSlotBit`, node reachability —
  if visited tracking uses a C array/list rather than a bitfield or hash, CWE-407 applies.
- **Lua scripting** — Redis embeds LuaJIT; Lua scripts can trigger Redis commands in
  dependency chains. Not directly graph-traversal, but Lua's own compiler (LuaJIT's
  `lj_ir.c`, `lj_opt_fold.c`) may have graph algorithm issues.
- **Module dependency** (`src/module.c`) — Redis modules declare dependencies. If
  topological sort of module load order uses list membership, that's CWE-407.

Redis is C; patterns to look for: linear array scan in cluster path computation.

### Memcached

Simpler architecture — primarily hash tables for key storage, slab allocator for memory.
Graph algorithms are minimal. **Low probability.** The `assoc.c` (hash table) and
`items.c` (LRU chains) don't do graph traversal. Skip for now unless scan reveals hits.

### Varnish Cache

Varnish uses VCL (Varnish Configuration Language) which is compiled to C and then loaded
as a shared library. The VCL compiler (`lib/libvcc/`) has:
- **AST construction and traversal** — VCL is parsed into an AST, then compiled.
  If the compiler uses list-based visited sets in tree/graph traversal, CWE-407 applies.
- **`vcc_compile.c`, `vcc_backend.c`** — backend (upstream) dependency tracking.
  If backends form a dependency graph (director chains), cycle detection may use lists.

Varnish's VCL compiler runs at config load time, not per-request — lower urgency.
But large Varnish deployments with complex VCL (hundreds of backends, subroutine chains)
could see slow reload times.

### nginx

nginx's architecture is event-driven with minimal graph structure. However:
- **`src/core/ngx_resolver.c`** — DNS resolver. Resolves chains of CNAMEs.
  CNAME chains are effectively a linked list, but cycle detection (detecting CNAME loops)
  uses a linear scan of the chain. For most cases this is bounded (max 8 CNAME hops),
  but the pattern is CWE-407 if implemented naively.
- **`src/http/ngx_http_upstream.c`** — Upstream group management. If upstream
  health-check state uses list-based membership, it's O(n) per check.
- **`src/http/ngx_http_rewrite_module.c`** — Rewrite rule chains. `break`/`last` flags
  terminate chains, so bounded. Low probability.

nginx is largely clean architecturally — it doesn't do complex graph computation at
runtime. Build-time scan still warranted.

### Apache2 (httpd)

- **`server/config.c`** — Module configuration merging. If module dependency graph
  uses `ap_array_make` (Apache's C array) for visited tracking, CWE-407 applies.
- **`modules/proxy/mod_proxy_balancer.c`** — Load balancer worker state. If worker
  health state uses linear scan, O(n) per check on every request.
- **`modules/mappers/mod_rewrite.c`** — Rewrite rule evaluation. Chain of rules with
  conditions — if rule graph uses list membership for cycle detection, CWE-407.
- **`server/request.c`** — Request handler chain. `ap_run_*` hooks traverse handler
  lists — linear by design but bounded.

### PHP-FPM / mod_php / CGI

These are PHP execution environments, not independent graph algorithm implementations.
They execute PHP code (which uses Zend Engine) and manage process pools. The process
pool management (`fpm/fpm_children.c`, `fpm/fpm_scoreboard.c`) uses simple arrays —
not graph algorithms. The interesting graph code is in Zend Engine itself (above).

### Web infrastructure scan backlog

Add to `tools/Makefile` and `tools/scans/`:

| Target | Language | Key files | Priority |
|--------|----------|-----------|----------|
| `php-zend` | C | `Zend/zend_cfg.c`, `zend_dfg.c`, `zend_ssa.c`, `ext/opcache/Optimizer/` | HIGH |
| `php-composer` | PHP | `src/Composer/DependencyResolver/` | MEDIUM |
| `redis` | C | `src/cluster.c`, `src/module.c` | MEDIUM |
| `varnish` | C | `lib/libvcc/vcc_compile.c`, `vcc_backend.c` | MEDIUM |
| `nginx` | C | `src/core/ngx_resolver.c`, `src/http/` | LOW |
| `apache2` | C | `server/config.c`, `modules/proxy/mod_proxy_balancer.c` | LOW |

---

## P2P and Anonymity Network Infrastructure

Scanned 2026-03-24. Six systems checked; one confirmed defect.

### Confirmed defect — Tor

**tor-0001** — `src/feature/nodelist/routerlist.c:2179`

`router_load_routers_from_string()` validates received descriptors against a `smartlist_t`
of requested fingerprints using `smartlist_contains_string()` — a `for`-loop strcmp scan —
inside a `SMARTLIST_FOREACH_BEGIN` over all received descriptors. Cost: O(R²), where R is
the number of router descriptors in the batch.

- R at a directory authority: ~7,000-8,000 (full relay consensus at startup)
- `smartlist_contains_string` confirmed O(n) linear scan at smartlist.c:97
- Same pattern in extrainfo path at line 2263-2295
- Fix: replace `smartlist_t *requested_fingerprints` with `digestmap_t *` — Tor's
  existing O(1) map, already used correctly in the same file at lines 2689, 2717

**Status:** UNPATCHED — ticket `tor-0001.md`

### CLEAN systems

| System | Language | Key finding |
|--------|----------|-------------|
| I2P Java router | Java | `tunnel/pool/TunnelPeerSelector.java` uses `Set<Hash>` (HashSet) throughout |
| libtorrent | C++ | `std::find` calls are assert-only or protocol-bounded (≤10 items per fast-set) |
| Transmission | C++ | Minimal `std::find`; no graph traversal hot paths |
| Kubo (go-ipfs / IPFS) | Go | Map-first idiom; no slice-backed visited sets in DAG traversal |
| Deluge | Python | List `.index()` calls are UI-only GTK operations |

### P2P scan backlog

| Target | Language | Key files | Priority |
|--------|----------|-----------|----------|
| `i2p.i2p router` | Java | `router/java/src/net/i2p/router/networkdb/` — KBucket/NetDB operations | MEDIUM |
| `libp2p-go` | Go | `routing/` — Kademlia DHT traversal | LOW |
| `zeromq` | C++ | `src/` — message routing graph | LOW |

---

## Eighth frontier: financial markets infrastructure

Financial markets are the highest-stakes environment in which this defect map operates.
The patches touch every layer of the financial stack — from the network routing that
carries market data, to the compilers that build trading systems, to the message brokers
that route orders, to the databases that hold positions. Each layer has its own blast
radius profile.

### Layer 1 — Network (OSPF in exchange co-location)

**frrouting-0002 is unpatched and directly affects financial market reliability.**

Stock exchanges, dark pools, and electronic trading venues operate in co-location
facilities where low-latency connectivity is the product. These facilities run OSPF
internally for routing between cabinets and switching layers. Every link failure —
a transceiver fault, a scheduled maintenance failover, a cable pull — triggers an
OSPF SPF recalculation.

With frrouting-0002 unpatched, SPF on a hub-and-spoke co-location topology (all racks
connected to a core switch layer — the dominant design) is O(V²) per topology change.
For a facility with 500 connected endpoints, that is ~125,000 comparisons per failover
event instead of ~500. OSPF convergence delay is directly proportional to the time
trading systems are unreachable during failover.

**The market impact:** If OSPF convergence takes longer than expected, trading systems
that rely on co-location connectivity experience unexpected latency spikes or brief
disconnection. For algorithmic trading systems with sub-millisecond latency requirements,
this is indistinguishable from a market data outage. Orders may be rejected, hedges may
fail to execute, and risk positions may be left unhedged during the convergence window.

Priority: patch frrouting-0002 and notify co-location facility operators (Equinix,
NYSE Mahwah, CME Aurora, CBOE Lenexa) as part of coordinated disclosure.

### Layer 2 — FIX protocol engines

The Financial Information eXchange (FIX) protocol is the message layer of every
electronic market. Every order, cancel, execution report, and market data update flows
through a FIX engine.

**QuickFIX/J** (Java) — the dominant open-source Java FIX engine. Used by brokers,
hedge funds, and exchanges globally. Compiled with javac → javac patches apply to every
QuickFIX/J build. Session management and routing logic involves graph traversal for
session dependency resolution.

**QuickFIX** (C++) — the C++ FIX engine. Compiled with GCC/Clang. LTO used in
production builds → llvm-0001 applies. The session graph and message routing logic are
candidates for their own CWE-407 scan; QuickFIX C++ has not been directly scanned.

**Recommendation:** Scan QuickFIX C++ `src/` for `std::find` / `std::vector::contains`
patterns in session graph and routing code. Given the age of the codebase (2000s-era)
and the language, probability of CWE-407 candidates is medium-high.

### Layer 3 — Order management and trading systems

**Java trading systems** — the majority of exchange-facing trading systems at major
financial institutions are built on Java. Order management systems (OMS), execution
management systems (EMS), and smart order routers all compile with javac. All benefit
directly from javac-0001 through javac-0005.

**Scala/Akka trading systems** — Akka is the dominant actor framework for high-throughput
Scala trading backends. Used at LMAX, Goldman Sachs (SecDB), Morgan Stanley, and
quantitative hedge funds. **scala3-0001 (O(n³) type inference) hits every Scala 3 trading
codebase directly.** A cubic constraint solver in the compiler means every build of a
type-heavy Akka or Cats Effect trading application was running at cubic cost. The patch
reduces this to linear.

**C++ HFT systems** — high-frequency trading firms (Virtu, Citadel Securities, Jane
Street, Two Sigma) build almost exclusively in C++ for sub-microsecond latency. All
benefit from llvm-0001 (LLVM LTO in release builds) and gcc-0001. The HFT build cycle
is aggressive — rebuilds happen frequently as strategies are updated. Faster LTO
directly reduces the time between strategy change and live deployment.

**Kotlin fintech backends** — kotlin-0001 (inheritance restriction checking) affects
every Kotlin financial services backend. Corda/R3 is the canonical example, but
Kotlin is now the default language at many fintech firms (Revolut, Monzo, N26,
Stripe's backend services).

### Layer 4 — Message brokers and event streaming

**Apache Kafka** — the dominant event streaming platform for financial data. Used at
every major exchange, bank, and trading venue for market data feeds, trade events, and
risk streams. Kafka is Java, compiled with javac. The Kafka broker's internal dependency
graph and topic partition assignment logic benefit from javac patches. Kafka Streams
(Scala/Java) benefits from both javac and scala3-0001.

**RabbitMQ** — Erlang-based message broker used heavily in financial messaging
infrastructure. Faster digraph ops (erlang-0001 patched, erlang-0002 FIXABLE-UPSTREAM)
improve exchange graph validation. **Throttle risk applies** — see Erlang ecosystem
section above. RabbitMQ at financial scale (stock exchanges, clearinghouses) must be
audited before the OTP patch is deployed.

**LMAX Disruptor** — Java ring buffer / event processing framework designed specifically
for financial low-latency systems. Used at LMAX Exchange and widely adopted in financial
middleware. Pure Java; compiled with javac. Benefits from inference patches in any
generic-heavy usage.

### Layer 5 — Risk and position databases

**PostgreSQL in financial analytics** — risk management systems, position databases,
P&L calculation engines, and regulatory reporting systems (MIFID II, Dodd-Frank) run
heavily on PostgreSQL. The five deferred CWE-407 defects in the PostgreSQL query planner
are directly material here:

- `postgresql-0002` (MERGE/UPDATE planning) — financial systems use MERGE heavily for
  upsert patterns in position and trade tables. Wide position tables (50–200 columns)
  and complex MERGE statements hit the O(W²×C²) defect directly.
- `postgresql-0003` (equivalence class matching) — analytical risk queries with many
  join predicates (risk factor joins, scenario analysis) hit the O(M×E) inner loop.
- `postgresql-0004` (join elimination) — self-join patterns common in slowly-changing
  dimension tables (instrument reference data, counterparty master) trigger this path.

PostgreSQL in financial infrastructure is one of the strongest arguments for prioritizing
the `nodeHash()` contribution. The query planner defects are not abstract — they affect
every complex analytical query against wide financial tables.

**TimescaleDB** — time-series extension to PostgreSQL, used for market data storage.
Inherits all five PostgreSQL planner defects. Time-series financial queries (OHLCV,
tick data, order book snapshots) against wide tables trigger the same paths.

### Layer 6 — TypeScript trading platforms

Modern trading platforms, broker portals, and market data dashboards are TypeScript-
heavy. Bloomberg's web terminal, Refinitiv Eikon Web, and the majority of broker
execution portals are TypeScript SPAs. ts-0001 through ts-0003 affect every TypeScript
trading frontend:

- **Developer experience:** VS Code language server performance on large trading
  platform codebases. Symbol resolution latency, auto-complete lag in complex
  type-parameterized components. Financial UI codebases are type-heavy by design
  (price types, instrument types, order state machines).
- **CI/CD build time:** TypeScript type-checking in CI for every trading platform
  frontend. Faster type-checking means faster deployment of trading UI changes.

**npm dependency resolution** (npm-0002, arborist patch) — every `npm install` for
trading platform frontends. Financial firms run npm install constantly in CI.

### Layer 7 — DeFi and on-chain financial systems

**solc-0001 (HIGH, unpatched)** — every Solidity contract compiled with `--via-ir`
or `--optimize` is affected. DeFi protocols — Uniswap, Aave, Compound, Curve,
MakerDAO — compile all production contracts through the Yul IR pipeline. For contracts
with deep internal function call graphs (lending protocols with complex liquidation
logic, DEX routers with many hop paths), the O(F×D²) Yul cycle detection is a real
compile-time cost. More critically: **solc is part of the security audit process**.
Every smart contract security audit involves multiple recompilations with different
optimization settings. A slow compiler increases audit costs and may compress the time
auditors spend on each compilation step.

### Cross-layer risk: deployment velocity in financial systems

Faster build pipelines mean faster deployment of fixes — and faster deployment of
mistakes. This is the dual-use concern for financial systems specifically:

**The upside:** A critical trading system bug discovered at market open can be
hotfixed and deployed faster. The window between discovery and remediation shrinks.
For financial systems where a bug can cost millions per minute, this is real value.

**The downside:** Financial systems have strict change management. Deployments go
through approval chains, pre-deployment testing, and regulatory notification for
certain changes. A faster build pipeline does not shorten the approval chain. The
risk is that development teams, experiencing faster builds, develop habits around
faster iteration that collide with the change management requirements. "It builds
faster so we can deploy faster" is not a valid rationale for bypassing change control.

**Mitigation:** Ensure that change management processes are decoupled from build
time. Faster CI should translate to more test coverage per deployment, not fewer
gates before production.

### Financial markets blast radius summary

| Layer | Systems affected | Patches | Risk profile |
|-------|-----------------|---------|--------------|
| Network routing | OSPF in co-location | frrouting-0002 (unpatched) | HIGH — live reliability |
| FIX engines | QuickFIX/J, QuickFIX C++ | javac, llvm-0001 | Medium — build pipeline |
| Trading systems | Java OMS/EMS, Scala/Akka, C++ HFT, Kotlin fintech | javac, scala3, llvm, kotlin | Low-Medium — faster builds |
| Message brokers | Kafka, RabbitMQ | javac, erlang | Medium — Erlang throttle risk |
| Risk databases | PostgreSQL, TimescaleDB | deferred ×5 | Medium — planning defects in prod |
| Trading UIs | TypeScript platforms | ts-0001..0003 | Low — build/dev experience |
| DeFi/on-chain | Solidity (Ethereum) | solc-0001 (unpatched) | Medium — audit pipeline |
| Build pipeline | All of the above | All patches | Dual-use — velocity is good and dangerous |

---

## Ninth frontier: game engine ecosystems

### Minecraft Java Edition — server-26.1

Minecraft Java Edition is the world's best-selling PC game and one of the most widely
deployed custom server ecosystems in existence. Hundreds of thousands of servers run
community-operated instances; the modded server ecosystem (Forge, Fabric, NeoForge)
adds thousands of additional mods per major version. The server jar is bytecode-only
(no source); analysis was performed via CFR decompiler on the extracted inner jar
(`META-INF/versions/26.1/server-26.1.jar`, 7,351 classes, server-26.1).

**Scan method:** CFR decompiler + Python bytecode string scan for `List/contains`
patterns across all 7,351 classes. Core graph utilities verified clean.

#### minecraft-0001 — DependencySorter.isCyclic (HIGH)

**File:** `net/minecraft/util/DependencySorter.java` (decompiled)
**Called from:** `net/minecraft/tags/TagLoader` — tag dependency resolution
**Trigger:** Every world load, every `/reload`, every `/datapack enable`
**Complexity:** O(E^D) worst case — exponential, no visited set in recursive DFS

`DependencySorter.isCyclic()` performs a recursive DFS to check whether adding a
dependency edge would create a cycle. No visited set. For a diamond dependency graph
of depth D, the number of node visits is 2^D:

```java
private static <K> boolean isCyclic(Multimap<K, K> directDependencies, K from, K to) {
    Collection dependencies = directDependencies.get(to);
    if (dependencies.contains(from)) {
        return true;
    }
    return dependencies.stream().anyMatch(
        dep -> DependencySorter.isCyclic(directDependencies, from, dep)
    );
}
```

**Impact:** Called from `TagLoader` — Minecraft's classification system (`#minecraft:logs`,
`#forge:ores/iron`, etc.). Tags can reference other tags. The dependency sort runs on
every world load and every `/reload`.

Vanilla Minecraft has hundreds of tags — tolerable. Large modpacks (Create, Applied
Energistics 2, Mekanism, Thermal Expansion) have thousands of cross-mod tag
dependencies with endemic diamond inheritance patterns. The "tag loading lag" reported
by modpack server operators — multi-second freezes on every server start and `/reload` —
is consistent with O(E^D) revisiting on diamond-shaped tag dependency graphs.

**Fix:** Add `Set<K> visited` parameter to track explored nodes. Per-call cost drops
from O(E^D) to O(E); total tag loading from O(E^D × E) to O(E²). Better fix: single
SCC pass over the complete graph after all edges are added, reducing total cost to
O(V+E).

**Disclosure path:** bugs.mojang.com (public bug tracker, "Performance" category)

#### minecraft-0002 — PistonStructureResolver.toPush (LOW)

**File:** `net/minecraft/world/level/block/piston/PistonStructureResolver.java`
**Pattern:** `this.toPush.contains(start)` — `toPush` is `ArrayList<BlockPos>`
**Complexity:** O(P²) — bounded at P≤12 by game design

Every piston activation resolves a push chain. Duplicate detection uses
`ArrayList.contains()` — linear scan. Minecraft hardcodes a maximum of 12 pushed
blocks per piston:

```java
if (blockCount + this.toPush.size() > 12) { return false; }
```

P≤12 caps the defect at 144 comparisons per piston. However, redstone contraptions
with many pistons firing simultaneously compound this: a 16×16 piston array at 20
TPS produces 737,280 list comparisons per second. Principle violation; low priority.

**Fix:** Parallel `HashSet<BlockPos>` for O(1) duplicate detection, same as javac-0001
(Tarjan stack → parallel set pattern).

#### Confirmed CLEAN

| Class | Why clean |
|-------|-----------|
| `util/Graph.depthFirstSearch` | Uses `Set<T>` for both `discovered` and `currentlyVisiting` — O(1) contains |
| `util/FeatureSorter` | Uses `TreeSet` for visited/onStack — O(log n), deliberate for deterministic ordering |
| `util/DependencySorter.visitDependenciesAndElement` | Uses `HashSet alreadyVisited` — O(1) |
| `world/level/lighting/DynamicGraphMinFixedPoint` | No list containers in bytecode |
| `world/level/chunk/status/ChunkDependencies` | No list containers in bytecode |

**Note on broader scan:** 34 classes had the `ArrayList + contains` bytecode signature.
All other hits are bounded inputs (pack selection: tens of packs), non-hot paths
(advancement layout, crash report categories), or operate on `HolderSet` (Minecraft's
own set wrapper, likely O(1)). `DependencySorter.isCyclic` is the only site where
unbounded graph traversal occurs without visited tracking.

#### Modding ecosystem blast radius

The Minecraft modded ecosystem amplifies minecraft-0001 specifically:

| Actor | Impact |
|-------|--------|
| **Vanilla server operators** | Hundreds of tags — tolerable; load time unnoticed |
| **Small modpack servers (50–200 mods)** | Thousands of tags — measurable lag on `/reload` |
| **Large modpack servers (Create, ATM, Omnifactory)** | Thousands of cross-mod diamond deps — multi-second freeze per reload |
| **Modpack developers** | `/reload` during development is slow; iteration cycle harmed |
| **Server hosting providers** | Restart time SLAs affected for large-modpack plans |

minecraft-0001 is a live performance defect affecting every large modpack server start
worldwide. The tag loading lag is user-visible and widely reported; the root cause has
not previously been identified.

### Mod source scan — Create, AE2, Mekanism

Three major open-source mods were independently scanned for CWE-407:

**Create mod — create-0001 (MEDIUM).** `TrackGraph.findDisconnectedGraphs()` uses
`ArrayList.remove(0)` as the BFS frontier queue. `ArrayList.remove(0)` is O(n) — the
backing array shifts all remaining elements on every dequeue. O(V²) BFS instead of
O(V+E). Trigger: every track removal event. Fix: `ArrayDeque.removeFirst()`.

**Applied Energistics 2 — CLEAN.** `GridNode.java` uses `ArrayDeque` + integer
generation counter for visited tracking — O(1). `PathingService.java` uses `HashSet`
in loop — O(1).

**Mekanism — CLEAN.** `OrphanPathFinder` uses `ObjectOpenHashSet<BlockPos>` (fastutil)
+ `Deque<BlockPos>` — both O(1). Written with performance awareness.

| Mod | Defect | Severity | Trigger |
|-----|--------|----------|---------|
| Minecraft (all mods) | minecraft-0001 `DependencySorter.isCyclic` | HIGH/EXPONENTIAL | Every world load |
| Minecraft (all mods) | minecraft-0002 `PistonStructureResolver` | LOW | Piston activation |
| Create mod | create-0001 `TrackGraph.findDisconnectedGraphs` | MEDIUM | Track removal |
| AE2 | — | CLEAN | — |
| Mekanism | — | CLEAN | — |