@williamzujkowski/live-coding-music-mcp 3.0.0 → 4.1.0

package/dist/services/EventDensityProbe.js

@@ -0,0 +1,324 @@
+/**
+ * Decides whether a pattern query is safe to materialize (#360).
+ *
+ * ## What was wrong with the old guard
+ *
+ * `queryEvents` sampled the first 1e-4 of the requested range, extrapolated,
+ * and refused if the projection exceeded the cap. Measured, that guard has
+ * two independent holes:
+ *
+ *   s("bd*200000")        -> refused correctly
+ *   s("~ bd*200000")      -> OUT OF HEAP. One leading rest, and the probe
+ *                            sees nothing, projects nothing, refuses nothing.
+ *   s("[bd*99999]*99999") -> OUT OF HEAP. 1e-4 of a 10^10-event cycle is
+ *                            still 10^6 events; the probe dies before it
+ *                            can refuse.
+ *
+ * The second is the cheaper attack: it needs no density in the probe
+ * window at all, just ordinary musical notation.
+ *
+ * ## What this does instead
+ *
+ * Probes at several offsets spread across the range — so a rest at the
+ * start cannot hide anything — with the span grown geometrically from
+ * very small, so a dense pattern is caught while the sample is still
+ * tiny.
+ *
+ * ## The rule that makes it safe to extrapolate
+ *
+ * Measured, and it is the part that is easy to get wrong:
+ *
+ *   s("bd").slow(64)              at span 1e-12 -> 1 hap
+ *   stack(s("bd*4"), s("hh*8"))   at span 1e-12 -> 2 haps
+ *
+ * Naively projecting the first gives 10^12 events and refuses one of the
+ * most ordinary patterns there is. Events legitimately begin at t=0, and
+ * a window narrower than the true inter-onset interval ALWAYS
+ * over-projects. So a projection is only taken once the sample looks
+ * like a sample: at least `minDistinctOnsets` DISTINCT onset times
+ * inside the window. Distinct times, not hap count — a stack of N layers
+ * all starting together is one moment observed N times, and says nothing
+ * about density.
+ *
+ * (The old `PROBE_FRACTION = 1e-4` avoided this by arithmetic accident:
+ * one stray hap projects to 10^4, which happens to sit under the 50,000
+ * cap. Any smaller constant would have begun refusing real patterns.)
+ *
+ * ## Residual
+ *
+ * Sampling cannot be sound against arbitrary non-uniform density. Three
+ * shapes get past this, all of them contained since #307 as an error
+ * envelope and a respawned child rather than a dead server:
+ *
+ *   - density confined to a gap between every window, at every rung;
+ *   - a burst so dense that growing a window into it from an adjacent
+ *     rest is itself fatal, since growth is what discovers it;
+ *   - clustered events sitting exactly on a slice boundary, which the
+ *     phase offset deliberately steps away from.
+ *
+ * The last is the price of decorrelating from the cycle, and it is worth
+ * paying: without the offset, EVERY window on a multi-cycle range shares
+ * one phase, and a leading rest hides a pattern from all of them.
+ */
+/**
+ * Windows spread across the range. This is the guard's resolution: a
+ * dense region narrower than 1/OFFSETS of the requested range, sitting
+ * between two windows, is not seen.
+ *
+ * Six was tried and rejected — it misses density confined to the last
+ * EIGHTH, which is an ordinary thing for a pattern to do. Coverage is
+ * worth more here than the ~15ms it costs, because this runs on
+ * query_pattern_events, a diagnostic, and not on the write path.
+ */
+export const DEFAULT_OFFSETS = 8;
+export const DEFAULT_MIN_DISTINCT_ONSETS = 16;
+/**
+ * Onsets a window should hold before its density is believed.
+ *
+ * Small samples are unreliable in both directions and the error is not
+ * academic: measured, a 1e-3 window at t=0 of `s("bd*40000")` holds 80
+ * onsets where the cycle average is 40, so extrapolating from it
+ * projects 80,000 against a 50,000 cap and refuses a pattern that is
+ * comfortably legal. Growing the window until it holds a real sample
+ * costs a few thousand haps and removes the guesswork; it is a much
+ * better answer than a fudge factor on the threshold.
+ */
+export const DEFAULT_SAMPLE_TARGET = 2000;
+/**
+ * How far above the cap a SMALL sample must project before it is
+ * believed without growing the window.
+ *
+ * This is the escape hatch for the pathological case: at 10^10 events
+ * per cycle there is no growing the window — the next step up would
+ * materialize more than the heap holds. A sample projecting ten times
+ * the cap is not a borderline call, and waiting for a bigger one is how
+ * the old guard died.
+ */
+export const OBVIOUSLY_HUGE_MULTIPLE = 10;
+/**
+ * Three rungs, not five. Every rung costs a query at every offset, and a
+ * `queryArc` call has ~1.8 ms of fixed overhead regardless of how little
+ * it returns — so the ladder's length, not its contents, is what the
+ * ordinary sparse pattern pays for.
+ *
+ * The first rung has to be tiny or the probe is the thing that runs out
+ * of heap: 1e-4 of a 10^10-event cycle is a million events, which is how
+ * the guard this replaced died. From there, three orders of magnitude a
+ * step reaches a usable sample for anything the cap cares about.
+ */
+export const DEFAULT_LADDER = [1e-8, 1e-6, 1e-4, 1e-2];
+/**
+ * Fractional part of the golden ratio, used to offset each window within
+ * its own slice.
+ *
+ * Evenly spaced windows are not enough, and the failure is not subtle.
+ * `query_pattern_events` accepts up to 16 cycles; with 8 windows across
+ * 16 cycles every window starts on a whole cycle — the SAME PHASE, eight
+ * times. So a pattern with a leading rest hides from all of them:
+ *
+ *   s("~ bd*200000") over 0..1   -> refused
+ *   s("~ bd*200000") over 0..4   -> OUT OF HEAP
+ *   s("~ bd*200000") over 0..16  -> OUT OF HEAP
+ *
+ * Musical patterns repeat per cycle, so any window placement that is
+ * commensurate with the cycle samples one phase repeatedly and calls it
+ * coverage. An irrational step decorrelates the phases while keeping one
+ * window per slice, which pure golden-ratio placement would not — that
+ * leaves whole slices empty.
+ */
+const GOLDEN_FRACTION = 0.618033988749895;
+/**
+ * Hard ceiling on a probe window, in cycles, whatever the range.
+ *
+ * Ladder rungs are fractions of the requested range, and the range may be
+ * 16 cycles — so the top rung would be 0.16 cycles wide, and a window
+ * that starts in a rest and grows across the boundary into a dense region
+ * materializes in proportion to how far it reaches. Capping the absolute
+ * width bounds that exposure by 16x on the widest ranges without
+ * changing any verdict: density is per cycle, so a smaller window
+ * measures it just as well.
+ */
+const MAX_PROBE_CYCLES = 0.01;
+/**
+ * Decides whether querying `[start, end)` is safe.
+ *
+ * @param queryArc - The pattern's own query function
+ * @param start - Range start, in cycles
+ * @param end - Range end, in cycles
+ * @param options - Cap and probe shape
+ * @returns A refusal with its evidence, or permission to proceed
+ */
+export function probeEventDensity(queryArc, start, end, options) {
+    const span = end - start;
+    const offsets = options.offsets ?? DEFAULT_OFFSETS;
+    const minDistinct = options.minDistinctOnsets ?? DEFAULT_MIN_DISTINCT_ONSETS;
+    const sampleTarget = options.sampleTarget ?? DEFAULT_SAMPLE_TARGET;
+    const ladder = options.ladder ?? DEFAULT_LADDER;
+    const cap = options.maxEvents;
+    let queries = 0;
+    let observedOnsets = 0;
+    /** Onsets and window width accumulated across every BELIEVABLE sample. */
+    let sampledOnsets = 0;
+    let sampledSpan = 0;
+    /**
+     * Every onset seen in a decision sample, believable or not.
+     *
+     * Windows are disjoint, so this is a LOWER BOUND on the events inside
+     * the requested range — evidence rather than arithmetic. It is the only
+     * thing that catches clustered events: eight chords of ten thousand
+     * layers hold 80,000 events and describe no density at all, so the
+     * distinctness gate rejects every sample and the projection has nothing
+     * to work with. Counting still works. (Found by cross-model review
+     * after my first attempt refused to accumulate unbelievable samples and
+     * let exactly this through.)
+     */
+    let countedFloor = 0;
+    /** How many windows contributed anything. Used to sanity-check an empty result. */
+    let windowsWithOnsets = 0;
+    if (span <= 0 || offsets < 1) {
+        return { kind: 'proceed', queries, observedOnsets, windowsWithOnsets: 0 };
+    }
+    const step = span / offsets;
+    for (let i = 0; i < offsets; i++) {
+        // Where in this slice the window sits. Irrational in the cycle, so
+        // eight windows sample eight different phases rather than the same
+        // one eight times — see GOLDEN_FRACTION.
+        const phase = (i * GOLDEN_FRACTION) % 1;
+        for (let rung = 0; rung < ladder.length; rung++) {
+            const fraction = ladder[rung];
+            const probeSpan = Math.min(span * fraction, MAX_PROBE_CYCLES);
+            if (probeSpan <= 0)
+                continue;
+            // Slid within its own slice, and never so far that the window runs
+            // past the slice's end — which would put the last one past `end`.
+            const windowStart = start + i * step + phase * (step - probeSpan);
+            // Whether there is anywhere left to grow. The largest sample this
+            // offset can produce still has to be USED — discarding it because
+            // it fell short of the target is how a pattern at 120,000 events
+            // per cycle walked straight past the guard.
+            const canGrow = rung + 1 < ladder.length;
+            const windowEnd = windowStart + probeSpan;
+            const haps = queryArc(windowStart, windowEnd);
+            queries++;
+            // Onsets only. A note held across the window began somewhere else
+            // and is not evidence of density here.
+            const onsets = [];
+            for (const hap of haps) {
+                const begin = hap.whole?.begin.valueOf();
+                if (begin === undefined)
+                    continue;
+                if (begin >= windowStart && begin < windowEnd)
+                    onsets.push(begin);
+            }
+            observedOnsets = Math.max(observedOnsets, onsets.length);
+            // Counted, not projected. Every one of these onsets is inside the
+            // requested range, so seeing cap-many in a single window proves the
+            // range holds at least that many — no extrapolation, and no
+            // distinctness gate to pass.
+            //
+            // This is the hole cross-model review (agy) found: a pattern with
+            // thousands of layers starting together clears the sample target
+            // while failing the distinctness gate, and fell out of the ladder
+            // without a verdict. Simultaneous events say nothing about DENSITY,
+            // but they are still events, and there can be too many of them.
+            if (onsets.length >= cap) {
+                return {
+                    kind: 'refuse',
+                    projected: onsets.length,
+                    sampleCount: onsets.length,
+                    spanFraction: probeSpan / span,
+                };
+            }
+            // From the actual window width, not the nominal rung: the two
+            // differ once MAX_PROBE_CYCLES bites, and using the rung would
+            // under-project by the same factor.
+            const projected = Math.round((onsets.length * span) / probeSpan);
+            const distinct = new Set(onsets).size;
+            // Enough separate moments for the count to describe a rate rather
+            // than a chord. See minDistinctOnsets.
+            const believable = distinct >= minDistinct;
+            // Obviously huge: refuse on a small sample rather than grow the
+            // window into the heap. Only this shortcut acts on one window;
+            // everything nearer the cap is decided in aggregate below.
+            if (believable && projected > cap * OBVIOUSLY_HUGE_MULTIPLE) {
+                return {
+                    kind: 'refuse',
+                    projected,
+                    sampleCount: onsets.length,
+                    // The width actually queried, which is not the rung once
+                    // MAX_PROBE_CYCLES clamps it.
+                    spanFraction: probeSpan / span,
+                };
+            }
+            // A sample big enough to trust — or the biggest this offset will
+            // ever produce. Either way, record it and move to the next offset.
+            if (onsets.length >= sampleTarget || !canGrow) {
+                countedFloor += onsets.length;
+                if (onsets.length > 0)
+                    windowsWithOnsets++;
+                if (countedFloor > cap) {
+                    return {
+                        kind: 'refuse',
+                        projected: countedFloor,
+                        sampleCount: countedFloor,
+                        spanFraction: probeSpan / span,
+                    };
+                }
+                // EVERY window that reaches here counts, distinct or not.
+                //
+                // The distinctness gate exists to stop a rate being extrapolated
+                // from a handful of stacked events — `s("bd").slow(64)` returns
+                // one hap at any span, and projecting from it refused ordinary
+                // patterns (#360 rule 1). It was never meant to discard a window
+                // that is *full*: a window holding `sampleTarget` events is
+                // evidence of density whatever their onset spelling.
+                //
+                // Discarding it let a pattern through that the probe exists to
+                // refuse. Measured, 150,000 onset instants x 250 simultaneous
+                // layers — 37,500,000 events in one cycle:
+                //
+                //   before: {"kind":"proceed","observedOnsets":3750,
+                //            "windowsWithOnsets":8}
+                //
+                // Every one of the eight windows saw the density and none of
+                // them was allowed to say so: each cleared `sampleTarget` (so
+                // the ladder stopped) but failed `minDistinct` (so it was not
+                // recorded), and `countedFloor` stayed under the cap because
+                // eight small windows hold few events in absolute terms (#460).
+                //
+                // Unconditional, and it has to say so. The fix for that wrote
+                // `if (believable || windowIsFull || !canGrow)` — inside a
+                // branch already guarded by `onsets.length >= sampleTarget ||
+                // !canGrow`, which implies it. The condition was always true,
+                // so `believable` had no effect here and the code described a
+                // gate it did not have. Reaching this line means the window is
+                // full or maximal; both count, so there is nothing to test
+                // (#464). `believable` still gates the obviously-huge shortcut
+                // above, which is the place the distinctness rule belongs.
+                sampledOnsets += onsets.length;
+                sampledSpan += probeSpan;
+                break;
+            }
+        }
+    }
+    // Averaged across every window, not taken from the densest one.
+    //
+    // The max is right for the obvious cases and wrong for the borderline:
+    // measured, a window at t=0 of `s("bd*40000")` sees twice the cycle's
+    // average density, so projecting from it gives 80,000 against a 50,000
+    // cap and refuses a pattern holding 40,001 events. Averaging over
+    // windows spread across the range is what makes the near-cap verdict
+    // trustworthy; the shortcut above still handles anything wildly over.
+    if (sampledSpan > 0) {
+        const projected = Math.round((sampledOnsets * span) / sampledSpan);
+        if (projected > cap) {
+            return {
+                kind: 'refuse',
+                projected,
+                sampleCount: sampledOnsets,
+                spanFraction: sampledSpan / span,
+            };
+        }
+    }
+    return { kind: 'proceed', queries, observedOnsets, windowsWithOnsets };
+}