feat(transform): live-α charsPerToken closes not_profitable max-savings leak

The break-even gate (isCompressionProfitable) used a hard-coded
CHARS_PER_TOKEN=4 (Anthropic's English avg). Live Claude Code traffic
has α≈0.5-0.9 (1.1-2 chars/tok) — JSON-dense tool docs, structured
CLAUDE.md slabs, etc. Off-by ~3.5× on the text side at α=0.88.

Production trace (events.jsonl, 2026-05-19): 200+ requests with system
slabs 113K-170K chars hit reason="not_profitable" because the gate's
textTokensEq = len/4 was a fraction of the row-aware image-token cost.
Example event: slab=169_632 chars, cache_create_tokens=148_891. Gate's
textEq=42K rejected vs imgCost=127K. With live α (1.14 ch/tok),
textEq=149K → ACCEPT. We were leaving real cache_create money on the
table on every one of those calls.

- TransformOptions gains charsPerToken?: number. Default 4 (back-compat).
- isCompressionProfitable() takes charsPerToken as a 5th optional param;
  defensive clamp on ≤0 / NaN / Infinity falls back to CHARS_PER_TOKEN.
- All 4 internal gate call sites (static slab + reminder + tool_result
  string + tool_result array) thread o.charsPerToken through.
- ProxyConfig.transform now also accepts a function: () => TransformOptions.
  Resolved per-request inside createProxy so the host can inject dynamic
  values. Static-object form unchanged for Workers / tests.
- node.ts wraps the static transform options in a closure that asks
  dashboardState.fitCosts() each request and passes the live
  chars_per_token. Fit-null falls through to the baked-in default — no
  chicken-and-egg deadlock since normal-sized slabs still feed the fit
  ring even before the huge-slab cases get unblocked.
- 4 new gate tests: stale-default rejects the 170K newline-heavy slab;
  live α=1.14 flips the same slab to profitable; defensive clamp on
  bogus charsPerToken; end-to-end transformRequest passes through.

Tests: 269 → 273 (+4). Typecheck + build clean.

src/core/proxy.ts

@@ -15,8 +15,12 @@ export interface ProxyConfig {
   upstream?: string;
   /** Override or supply an API key. If unset, we forward whatever the client sent. */
   apiKey?: string;
-  /** Per-request transform options. */
-  transform?: TransformOptions;
+  /** Per-request transform options. Pass a function when the host wants to
+   *  inject DYNAMIC values per request (e.g. live empirical `charsPerToken`
+   *  from the dashboard's converging fit) — the proxy invokes it once per
+   *  /v1/messages POST. Static object form is used by the Workers host and
+   *  tests that don't need dynamic state. */
+  transform?: TransformOptions | (() => TransformOptions);
   /** Called after every request — useful for logging / metrics in the host. */
   onRequest?: (event: ProxyEvent) => void | Promise<void>;
 }
@@ -321,7 +325,12 @@ export function createProxy(config: ProxyConfig = {}) {
     if (isMessages) {
       const bodyIn = new Uint8Array(await req.arrayBuffer());
       try {
-        const r = await transformRequest(bodyIn, config.transform);
+        // Resolve transform options per-request when the host passed a
+        // function — lets dashboardState.fitCosts()'s live α flow into
+        // isCompressionProfitable on every call.
+        const transformOpts =
+          typeof config.transform === 'function' ? config.transform() : config.transform;
+        const r = await transformRequest(bodyIn, transformOpts);
         // Cast: TS narrows Uint8Array<ArrayBufferLike> away from BodyInit, but
         // it's a valid body and we never use SharedArrayBuffer.
         bodyOut = r.body as unknown as BodyInit;

src/core/transform.ts

@@ -84,6 +84,19 @@ export interface TransformOptions {
    *  off until a smoke test against the real slab confirms ordering.
    *  Auto-clamped if the resulting canvas would exceed 1568 px wide. */
   multiCol?: number;
+  /** Chars-per-token assumption used by `isCompressionProfitable()`. Default
+   *  4 (Anthropic's published English-text average ≈ α=0.25). The host can
+   *  inject the live empirical `chars_per_token` from `dashboardState.
+   *  fitCosts()` per-request — content with dense tokenization (JSON-heavy,
+   *  tool definitions, Claude Code's CLAUDE.md slabs at α≈0.5-0.9) makes
+   *  the stale 4 chars/tok wildly under-estimate text-token equivalence,
+   *  and the gate then rejects compressions that would actually save real
+   *  tokens. With the live value plumbed in, the gate auto-tightens or
+   *  loosens to match the model's actual tokenization rate.
+   *  Production trace 2026-05-19: 200+ events with slabs 113K-170K chars
+   *  hit `not_profitable` because the gate used 4 chars/tok while live α
+   *  was ≈0.88 (1.14 chars/tok) — gate was off by ~3.5× on the text side. */
+  charsPerToken?: number;
 }
 
 const DEFAULTS: Required<TransformOptions> = {
@@ -118,6 +131,10 @@ const DEFAULTS: Required<TransformOptions> = {
   compressHistory: false,
   historyKeepTail: 4,
   historyMinPrefix: 10,
+  // English ~4 chars/tok default (= the CHARS_PER_TOKEN constant declared
+  // later in this file — kept as a literal here to avoid forward-reference).
+  // Host overrides per-request when the dashboard's live fit has converged.
+  charsPerToken: 4,
   // R2 multi-column ON (2 cols) — at single-col the break-even gate
   // correctly rejects compression on real tool-doc-shaped slabs (~38 chars/
   // row → ~29 imgs vs 39k text tokens → net loss). Two columns packs ~2×
@@ -204,6 +221,12 @@ export function isCompressionProfitable(
   cols: number = DEFAULTS.cols,
   imageCountCap?: number,
   numCols: number = 1,
+  /** Chars-per-token assumption for the text side of the break-even math.
+   *  Default 4 (Anthropic's English-text average). Host can pass the live
+   *  empirical `chars_per_token` from dashboardState.fitCosts() so the gate
+   *  auto-tunes to the model's actual tokenization rate. Lower values =
+   *  more profitable text compressions (each char buys more tokens back). */
+  charsPerToken: number = CHARS_PER_TOKEN,
 ): boolean {
   const n = Math.max(1, numCols | 0);
   let estImages: number;
@@ -225,8 +248,14 @@ export function isCompressionProfitable(
   if (imageCountCap !== undefined && imageCountCap > 0) {
     estImages = Math.min(estImages, imageCountCap);
   }
+  // Defensive clamp: a corrupt or pathological charsPerToken (≤0 / NaN)
+  // would either crash or give a misleading-true. Fall back to the
+  // baked-in default in that case.
+  const cpt = Number.isFinite(charsPerToken) && charsPerToken > 0
+    ? charsPerToken
+    : CHARS_PER_TOKEN;
   const imageTokensCost = estImages * TOKENS_PER_IMAGE;
-  const textTokensEquivalent = textLen / CHARS_PER_TOKEN;
+  const textTokensEquivalent = textLen / cpt;
   return imageTokensCost < textTokensEquivalent;
 }
 
@@ -1241,7 +1270,7 @@ export async function transformRequest(
     Math.max(1, (o.multiCol | 0) || 1),
     Math.max(1, maxFittingCols(o.cols)),
   );
-  if (!isCompressionProfitable(combined, o.cols, undefined, numCols)) {
+  if (!isCompressionProfitable(combined, o.cols, undefined, numCols, o.charsPerToken)) {
     info.reason = `not_profitable (slab=${combined.length} chars)`;
     bumpPassthrough(info, 'not_profitable');
     info.outgoingTextChars = countOutgoingTextChars(req);
@@ -1345,7 +1374,7 @@ export async function transformRequest(
             continue;
           }
           const reminderText = (blk as TextBlock).text;
-          if (!isCompressionProfitable(reminderText, o.cols, undefined, numCols)) {
+          if (!isCompressionProfitable(reminderText, o.cols, undefined, numCols, o.charsPerToken)) {
             // Above threshold but image cost ≥ text cost. Net loss to compress.
             bumpPassthrough(info, 'not_profitable');
             processedExisting.push(blk);
@@ -1406,7 +1435,7 @@ export async function transformRequest(
               if (inner.length < o.minToolResultChars) {
                 bumpPassthrough(info, 'below_threshold');
                 rewritten.push(blk);
-              } else if (!isCompressionProfitable(inner, o.cols, o.maxImagesPerToolResult, numCols)) {
+              } else if (!isCompressionProfitable(inner, o.cols, o.maxImagesPerToolResult, numCols, o.charsPerToken)) {
                 bumpPassthrough(info, 'not_profitable');
                 rewritten.push(blk);
               } else {
@@ -1450,7 +1479,7 @@ export async function transformRequest(
                   newInner.push(ib as TextBlock | ImageBlock);
                   continue;
                 }
-                if (!isCompressionProfitable(innerText, o.cols, o.maxImagesPerToolResult, numCols)) {
+                if (!isCompressionProfitable(innerText, o.cols, o.maxImagesPerToolResult, numCols, o.charsPerToken)) {
                   bumpPassthrough(info, 'not_profitable');
                   newInner.push(ib as TextBlock | ImageBlock);
                   continue;

src/node.ts

@@ -515,7 +515,21 @@ async function main(): Promise<void> {
 
   const config: ProxyConfig = {
     upstream: opts.upstream,
-    transform,
+    // Resolve transform options per request so the live empirical
+    // chars/token from dashboardState.fitCosts() flows into the
+    // isCompressionProfitable gate. With α at the stale default of 0.25
+    // (4 chars/tok), the gate rejects slabs as small as ~138K chars
+    // because the row-aware image-count estimate exceeds the textual
+    // token equivalence. Live α on Claude Code traffic runs ≈0.5-0.9
+    // — once the fit has 3+ samples, the gate auto-tunes and accepts
+    // compressions that the stale assumption was leaving on the table.
+    // When fit is null, falls through to the baked-in DEFAULTS.charsPerToken.
+    transform: () => {
+      const fit = dashboard.fitCosts();
+      return fit && fit.chars_per_token > 0
+        ? { ...transform, charsPerToken: fit.chars_per_token }
+        : transform;
+    },
     onRequest: async (e) => {
       // Feed the dashboard BEFORE tracker.emit — toTrackEvent strips
       // info.firstImagePng, so capturing has to happen on the raw event.