fix(vol4): resolve cross-references, heading case, unreferenced labels, and bib standards

Author: profvjreddi

books/vol3/frontmatter/about.qmd

@@ -29,7 +29,7 @@ When language models are given memory, external tools, and the authority to exec
 
 The fundamental thesis of this book is that **the atomic unit of systems engineering has shifted from the served request to the Managed Trajectory**.
 
-### The Invariant Closure Principle
+### The invariant closure principle
 
 Why does this shift demand an entirely new systems stack? The answer lies in what we term **The Invariant Closure Principle**:
 
@@ -96,6 +96,6 @@ This book is the third volume in the *Machine Learning Systems* series. Each vol
 
 While the chapters are designed to build progressively, readers with specific technical objectives may prioritize tailored paths:
 
-*   **The Systems & Infrastructure Engineer Path:** Focus on the memory and execution runtime: Chapter 1 (Trajectories) $\rightarrow$ Chapter 2 (ACB) $\rightarrow$ Chapter 6 & 7 (Context & Radix Prefix Caching) $\rightarrow$ Chapter 10 & 11 (MCP Protocols & MicroVM Sandboxes) $\rightarrow$ Chapter 12 (Sagas) $\rightarrow$ Chapter 15 (Observability).
-*   **The AI Researcher & Algorithm Architect Path:** Focus on search, planning, and error dynamics: Chapter 1 (Trajectories) $\rightarrow$ Chapter 3 (State Graphs & DAGs) $\rightarrow$ Chapter 4 (Scheduling) $\rightarrow$ Chapter 5 (Compounding Error & Test-Time Compute) $\rightarrow$ Chapter 8 (Scratchpads & Isolation) $\rightarrow$ Chapter 13 (Trajectory Benchmarks).
-*   **The Enterprise AI Architect & Security Lead Path:** Focus on governance, isolation, and economics: Chapter 1 (Invariant Closure) $\rightarrow$ Chapter 2 (Capability Leases) $\rightarrow$ Chapter 10 (Typed Contracts) $\rightarrow$ Chapter 11 (Blast Radius & Sandboxing) $\rightarrow$ Chapter 14 (Replay Debugging) $\rightarrow$ Chapter 16 (Fleet Economics).
+*   **The Systems & Infrastructure Engineer Path:** Focus on the memory and execution runtime: Part I (Trajectories, ACB) $\rightarrow$ Part II (Context & Radix Prefix Caching) $\rightarrow$ Part III (MCP Protocols, MicroVM Sandboxes, Sagas) $\rightarrow$ Part IV (Observability).
+*   **The AI Researcher & Algorithm Architect Path:** Focus on search, planning, and error dynamics: Part I (Trajectories, State Graphs & DAGs, Scheduling) $\rightarrow$ Part II (Compounding Error & Test-Time Compute, Scratchpads & Isolation) $\rightarrow$ Part IV (Trajectory Benchmarks).
+*   **The Enterprise AI Architect & Security Lead Path:** Focus on governance, isolation, and economics: Part I (Invariant Closure, Capability Leases) $\rightarrow$ Part III (Typed Contracts, Blast Radius & Sandboxing) $\rightarrow$ Part IV (Replay Debugging, Fleet Economics).

books/vol3/frontmatter/author_note.qmd

@@ -28,7 +28,7 @@ In *Introduction to Machine Learning Systems*, we optimized the single node: bal
 
 Volume III confronts what happens when that assumption shatters.
 
-Autonomy demands long horizons, but long horizons destroy reliability. If a model is ninety-eight percent accurate on any individual step, an unverified fifty-step autonomous workflow will fail nearly two-thirds of the time. Systems engineering is the only discipline that arrests that compounding decay. We must learn to manage context not as a bottomless prompt buffer, but as a structured memory hierarchy; to treat tool execution not as text strings, but as isolated, sandboxed system calls; and to build distributed Sagas with deterministic rollbacks when non-deterministic actions fail in the wild.
+Autonomy demands long horizons, but long horizons destroy reliability. If a model is 98 percent accurate on any individual step, an unverified fifty-step autonomous workflow will fail nearly two-thirds of the time. Systems engineering is the only discipline that arrests that compounding decay. We must learn to manage context not as a bottomless prompt buffer, but as a structured memory hierarchy; to treat tool execution not as text strings, but as isolated, sandboxed system calls; and to build distributed Sagas with deterministic rollbacks when non-deterministic actions fail in the wild.
 
 I write these chapters as a student of a discipline in motion. Putting this book together is my own way—as an educator and researcher—of wrestling with where the field is heading, sifting through the daily noise of transient frameworks, and boiling down the durable first principles of **Horizon, State, and Authority** that will still govern computing a decade from now.
 

books/vol4/appendix/cross-disciplinary-lexicon.qmd

@@ -27,11 +27,13 @@ Because each discipline developed its own vocabulary over decades, engineers fro
 
 ## 2. False Friends: Words That Cause Catastrophic Misunderstandings
 
+Engineering teams building Physical AI systems are drawn from distinct engineering traditions—deep learning, classical robotics, embedded systems, and functional safety. Because these disciplines evolved independently, they frequently use the same vocabulary to denote fundamentally incompatible concepts. These "false friends" represent subtle linguistic traps where an assumption from one field, when carried uncritically into another, leads directly to catastrophic hardware failures.
+
 ### "Deterministic"
 * **In Deep Learning:** Setting `torch.manual_seed(42)` and `cudnn.deterministic = True` so the matrix multiplication forward pass produces bit-identical tensor values across offline runs.
 * **In Embedded & Real-Time Systems:** Guaranteeing that every periodic task completes in hardware time $t \le T_{\text{deadline}}$ with bounded jitter ($<1\,\mu\text{s}$), regardless of CPU cache misses, interrupt storms, or bus contention. *An algorithm that always outputs the same tensor 10 milliseconds too late is non-deterministic in physical systems.*
 
-### "Real-Time"
+### "Real-time"
 * **In Web & Cloud Software:** A web service or UI that updates within $100\text{--}200\text{ ms}$ ("real-time chat", "real-time streaming"), where occasional delays cause dropped frames or buffering spinners without danger.
 * **In Physical AI Systems:** Hard deadline determinism where missing a single $1.0\text{ ms}$ execution deadline ($1000\text{ Hz}$) produces an inverter current loop runaway, mechanical resonance, or irreversible collision.
 
@@ -59,20 +61,22 @@ Because each discipline developed its own vocabulary over decades, engineers fro
 
 ## 3. Practical Translation Guides for Students
 
-### If you are coming from Computer Science / Deep Learning:
+Navigating the multi-disciplinary landscape of Physical AI requires recognizing the implicit assumptions of your home field and actively translating your engineering instincts to match physical realities. The following translation guides provide actionable principles tailored to students and practitioners entering Physical AI from specific foundational backgrounds.
+
+### If you are coming from computer science / deep learning:
 
 1. **Forget `malloc()` on the fast loop:** You cannot allocate dynamic memory or resize lists inside a $1000\text{ Hz}$ control loop. Heap fragmentation causes random multi-millisecond pauses that crash physical robots.
 2. **Your model is an unprivileged proposer:** A neural network must never have direct write access to motor registers. It proposes intent; a deterministic mathematical filter running on bare metal permits, clamps, or rejects the command.
 3. **Loss $= 0$ in simulation does not mean it works:** Physics engines approximate contact dynamics as linear complementarity problems (LCP). Real contact friction is non-smooth, stiction-dominated, and thermally variable.
 4. **Bandwidth matters more than FLOPS:** Streaming multi-gigabyte neural weights from LPDDR5 DRAM on edge SoCs creates a memory wall ($t_{\text{stream}} = M_{\text{weights}} / B_{\text{mem}}$) that limits inference cadence, regardless of how many TFLOPS the NPU matrix core claims.
 
-### If you are coming from Mechanical Engineering / Control Theory:
+### If you are coming from mechanical engineering / control theory:
 
 1. **Uncertainty is multimodal:** Classical Kalman filters assume unimodal Gaussian noise. Perception models in the real world face discrete semantic hypotheses (e.g., an object is either in front of or behind a partition), requiring multimodal policy representations (Diffusion / Mixture Density Networks).
 2. **You cannot write down the world dynamics $f(\mathbf{x})$ for an unconstrained kitchen or highway:** Analytical ODEs work for steel linkages in a factory cell, but fail when interacting with deformable objects, varying lighting, and open-world human behavior. Deep models provide the semantic perception front-end that classical control lacks.
 3. **Control Barrier Functions are the bridge:** You do not have to discard control theory when using deep learning. Instead, treat the learned policy as a nominal reference generator $\mathbf{u}_{\text{ref}}$ and use a Control Barrier Function (CBF-QP) as an active safety filter that minimally adjusts $\mathbf{u}$ only when boundary invariants are threatened.
 
-### If you are coming from Embedded Systems / Electrical Engineering:
+### If you are coming from embedded systems / electrical engineering:
 
 1. **AI compute is dynamic and stochastic:** You cannot treat neural inference like a fixed-cycle DSP FIR filter. Execution time varies based on image content, caching, and thermal throttling. You must design asynchronous two-speed architectures with expiring leases to absorb timing jitter safely.
 2. **Data provenance is as critical as voltage integrity:** Logging teleoperation data without tracking operator response delays ($340\text{ ms}$) or safety governor clamps poisons training pipelines via covariate shift.

books/vol4/appendix/heterogeneous-soc-reference.qmd

@@ -25,7 +25,7 @@ In an embodied machine, cognitive deliberation and mechanical actuation operate
 
 ## Shared Memory RPMSG Message Contract
 
-The application processor transmits timestamped intent proposals to the real-time safety controller over non-blocking shared SRAM mailboxes using the Remote Processor Messaging (RPMSG) protocol. The shared message schema (`PhysicalAgent_IntentProposal_t`) enforces a packed binary layout across cores:
+The application processor transmits timestamped intent proposals to the real-time safety controller over non-blocking shared SRAM mailboxes using the Remote Processor Messaging (RPMSG) protocol. The shared message schema (`PhysicalAgent_IntentProposal_t`), detailed in @tbl-ipc-intent-contract, enforces a packed binary layout across cores:
 
 | Field Name              | Type       | Size     | Physical Units & Description                                              |
 |:------------------------|:-----------|:---------|:--------------------------------------------------------------------------|
@@ -38,7 +38,7 @@ The application processor transmits timestamped intent proposals to the real-tim
 | `model_confidence`      | `float`    | 4 bytes  | Epistemic model confidence score $\gamma_{\text{conf}} \in [0.0, 1.0]$.   |
 | `crc32`                 | `uint32_t` | 4 bytes  | IEEE 802.3 32-bit CRC payload integrity checksum.                         |
 
-: Physical AI Dual-Brain IPC Intent Proposal Schema. Binary layout of the packed shared-memory RPMSG payload. {#tbl-ipc-intent-contract}
+: **Physical AI Dual-Brain IPC Intent Proposal Schema**: Binary layout of the packed shared-memory RPMSG payload. {#tbl-ipc-intent-contract}
 
 ## Hardware Safety Watchdog Protocol