LithosAnanake/docs/formal/scraps/architecture/PHYSICS_CONTROL_SYSTEM_DESIGN.tex

%% SCRAP: architecture/PHYSICS_CONTROL_SYSTEM_DESIGN
%% SOURCE: docs/working/architecture/PHYSICS_CONTROL_SYSTEM_DESIGN.adoc
%% STATUS: WORKING
%% FITS: dev-guide/app-physics
%% EDITORIAL: lifted — prose rewritten to press voice
%% PATENT: This scrap describes adaptive scheduling and memory-management
%% mechanisms driven by runtime physics observables. Several passages read as
%% claim-like ("physics metrics drive cache-tier placement", "temperature-weighted
%% eviction"). Flagged inline with %% PATENT: markers. No claims are drafted here;
%% forward to patent counsel before any external release.

\section{From Observability to Control}

The physics engine answers the question of what the system is doing. Scheduling
and memory management answer the question of what it should do differently. The
physics engine supplies observability --- entropy slope, temperature trend, p99
latency. Scheduling and memory management supply control: execution placement and
data placement respectively, joined by feedback-control loops.

Three subsystems consume physics signals. The scheduling system governs word
priority, core affinity, preemption hints, latency SLOs, and throttling. The
memory-management system governs cache tier, block placement, garbage-collection
triggers, and eviction policy. Between them, feedback loops close observation onto
action.

\subsection{The Scheduling System}

StarForth executes on top of a host scheduler (L4Re or Linux). The VM runs
largely on a single thread, cannot directly reorder word execution within a
definition, and influences the host only through priority hints
(\texttt{setpriority}, \texttt{sched\_setparam}); the return stack dictates
nesting order.

%% PATENT: claim-like phrasing follows.
Within these limits, physics can still control scheduling. At the OS level it sets
per-word priority (hotter words receive lower nice values), CPU affinity (hot,
cache-sensitive words pinned together; interfering words separated), and preemption
hints (hot, rapidly growing words made non-preemptible for batching).

\begin{lstlisting}[language=C]
typedef struct {
    int os_priority;       // nice level or sched_priority
    int scheduling_class;  // SCHED_OTHER, SCHED_FIFO, SCHED_RR
    int cpu_affinity_mask; // which cores allowed
} sched_hint_t;

// hotter = lower nice = higher priority
word->sched_hint.os_priority = (word->temperature_q8 > 0x8000) ? 5 : 10;
\end{lstlisting}

Even without host control, a VM-level scheduler can respect physics. Inside
\texttt{execute\_colon\_word()}, hot words on a critical path can be executed with
affinity or batched with related words, while cold words are deferred or batched
for idle time. Adaptive batching groups words that physics observes always run
together --- \texttt{IF}/\texttt{THEN}, for instance --- to minimize cache misses.

A latency-SLO mechanism times each execution against a per-word limit and, on
violation, adjusts knobs: it boosts priority, lowers the stack-depth limit, and
shortens the batch timeout, logging the breach.

\subsection{The Memory-Management System}

StarForth uses a single 5~MB dictionary heap (statically allocated, no
fragmentation), block storage of 1024 blocks of 1024 bytes, and two 1024-cell
stacks. There is no explicit memory manager; allocation is implicit in the
dictionary.

%% PATENT: claim-like phrasing follows.
Physics can drive placement across a logical memory hierarchy. A tier is chosen
from word temperature and entropy slope:

\begin{lstlisting}[language=C]
typedef enum {
    MEM_TIER_L1, MEM_TIER_L2, MEM_TIER_L3,
    MEM_TIER_DRAM, MEM_TIER_BLOCK
} mem_tier_t;

mem_tier_t choose_tier(DictEntry *word) {
    if (word->physics.temperature_q8 > 0xE000) return MEM_TIER_L1;
    if (word->physics.temperature_q8 > 0x8000) return MEM_TIER_L2;
    if (word->physics.temperature_q8 > 0x4000) return MEM_TIER_L3;
    if (word->physics.entropy_slope > 0)       return MEM_TIER_DRAM;
    return MEM_TIER_BLOCK;  // cold: evict to block storage
}
\end{lstlisting}

Cache coloring uses the learned call graph: words that frequently co-execute are
placed in different cache sets to avoid conflict. Garbage collection is triggered
not by a simple capacity threshold but by memory pressure (drawn from the host PSI
snapshot) combined with the presence of cold, stale words to reclaim. Eviction
replaces pure LRU with a temperature-weighted score --- age multiplied by inverse
temperature --- so a candidate is hot only when both old and cold:

\begin{lstlisting}[language=C]
uint64_t score = (age / 1000000) * ((0xFFFF - temperature_q8) + 1);
// older + colder => higher score => evicted first
\end{lstlisting}

Block storage absorbs the coldest words. A word spills to disk when it is cold, has
not run for several minutes, sits outside any SLO-critical path, and memory
pressure is high; it is lazily reloaded on next use, with its physics re-touched on
access.

\subsection{Integration Points}

Physics observes (entropy, temperature, recency, latency, call graph, stack depth,
error rate, memory pressure) and feeds the scheduler, which decides priority,
affinity, batch group, preemption, SLO, throttling, and stack limits. Execution
respects those hints and generates new metrics, which feed the memory manager's
decisions about cache tier, block placement, GC trigger, eviction, spilling, and
prefetch. Placement then affects the next iteration's execution, closing the loop.

Two narrow APIs connect the components. The physics-to-scheduling interface
(\texttt{sched\_control\_t}) carries priority, affinity mask, batch group,
preemptibility, latency limit, and stack limit; \texttt{sched\_update\_word()}
applies adjustments and \texttt{sched\_report\_execution()} reports observed
latency. The physics-to-memory interface (\texttt{mem\_control\_t}) carries
preferred tier, cache-set preference, spill-age threshold, evictability, and
eviction priority. Scheduler and memory manager also coordinate directly: a batch
decision notifies the memory manager so it can cache-color the group, and an
eviction of a critical-path word warns the scheduler before disabling that word's
fast path.

\subsection{Implementation Phases}

\begin{itemize}
  \item \textbf{Phase 1 --- Instrumentation (current).} Physics observes and
        publishes metrics; scheduling and memory read them advisorily.
  \item \textbf{Phase 2 --- Weak Control.} Scheduling adjusts OS priority hints;
        memory becomes cache-tier aware; still advisory.
  \item \textbf{Phase 3 --- Strong Control.} Scheduling enforces the latency SLO;
        memory enforces GC and eviction; the feedback loop is closed.
  \item \textbf{Phase 4 --- ML Integration.} A model predicts latency, memory, and
        scheduling outcomes; governance approves knob adjustments.
\end{itemize}

\subsection{Worked Examples}

\paragraph{Hot word going critical.} An I/O word heats during a file scan. Physics
reports rising temperature and entropy slope and alerts both subsystems. Scheduling
boosts the word's priority, enables I/O batching, caps recursion, and sets a 10~$\mu$s
SLO. Memory promotes the word to L2, prefetches related words
(\texttt{BLOCK\_WRITE}, \texttt{BUFFER}, \texttt{UPDATE}), and protects it from
eviction. Execution then meets the SLO at p99~$\approx$~8.5~$\mu$s while the
temperature stabilizes, after which both subsystems can act more aggressively on the
now-learned pattern.

\paragraph{Cold word emergency eviction.} Under 95\% dictionary use and 96\% cgroup
memory, GC triggers. Physics scores eviction candidates by age times inverse
temperature, selecting the coldest, oldest word first while protecting a recent hot
word on the critical path. The chosen word spills to block storage, is marked
\texttt{WORD\_SPILLED}, and reclaims space. Much later, when the user invokes it
again, the executor detects the flag, reloads it from storage in roughly 50~$\mu$s,
runs it, and lets physics re-learn its pattern.

\subsection{Critical Design Decisions}

\begin{itemize}
  \item \textbf{Scheduling granularity} --- per-word priority, coarse word
        batches, or a hybrid that starts fine-grained and learns optimal batches.
        Recommendation: hybrid.
  \item \textbf{Memory placement accuracy} --- static placement, dynamic
        migration, OS-hint delegation, or logical tiers with no physical movement.
        Recommendation: logical tiers as scheduling hints.
  \item \textbf{GC versus spilling} --- always compact, always archive, or a
        governance-driven per-word policy. Recommendation: governance-driven.
\end{itemize}

\subsection{Open Questions}

The design leaves several questions for resolution: whether the VM should implement
its own scheduler or defer entirely to the host; whether dictionary words can be
physically relocated without breaking pointers, or only hinted logically; whether
block storage serves persistence or active-memory overflow; how scheduling and
memory decisions should interact with L4Re IPC deadlines and capability delegation;
who owns eviction and spilling policy; and whether the added control overhead
(roughly 100~ns atop the existing 200--300~ns per event) is acceptable for target
latencies.

%% TODO(bob): Confirm acceptable per-event control overhead budget against the
%% current latency targets before committing to closed-loop enforcement in Phase 3.