diff --git a/src/core/contextual-retrieval-service.ts b/src/core/contextual-retrieval-service.ts
index 0e22b7746..a65c8e074 100644
--- a/src/core/contextual-retrieval-service.ts
+++ b/src/core/contextual-retrieval-service.ts
@@ -54,6 +54,7 @@ import {
 import {
   generatePerChunkSynopsis,
   SYNOPSIS_PROMPT_VERSION,
+  SYNOPSIS_DOC_MAX_CHARS,
   type GeneratePerChunkSynopsisResult,
 } from './page-summary.ts';
 import {
@@ -103,8 +104,17 @@ function getEmbeddingModelTag(): string {
 export function computeCorpusGeneration(args: {
   crMode: CRMode;
   haikuModel: string;
+  /**
+   * Resolved `SYNOPSIS_DOC_MAX_CHARS` for per_chunk_synopsis runs. When
+   * present, folded into the hash so changes to
+   * `GBRAIN_SYNOPSIS_DOC_MAX_CHARS` invalidate the prior cache cleanly.
+   * Omit for `crMode !== 'per_chunk_synopsis'` — title / none modes
+   * don't consult the cap and the field stays out of the hash for
+   * back-compat with pre-cap embeddings.
+   */
+  synopsisDocMaxChars?: number;
 }): string {
-  return createHash('sha256')
+  const h = createHash('sha256')
     .update(args.crMode)
     .update('|')
     .update(String(SYNOPSIS_PROMPT_VERSION))
@@ -113,9 +123,11 @@ export function computeCorpusGeneration(args: {
     .update('|')
     .update(String(TITLE_WRAPPER_VERSION))
     .update('|')
-    .update(getEmbeddingModelTag())
-    .digest('hex')
-    .slice(0, 16);
+    .update(getEmbeddingModelTag());
+  if (args.synopsisDocMaxChars !== undefined) {
+    h.update('|doc_cap=').update(String(args.synopsisDocMaxChars));
+  }
+  return h.digest('hex').slice(0, 16);
 }
 
 /**
@@ -253,7 +265,11 @@ export async function reembedPageWithContextualRetrieval(
       args.pageSlug,
       args.sourceId,
       resolution.mode,
-      computeCorpusGeneration({ crMode: resolution.mode, haikuModel: args.haikuModel ?? DEFAULT_HAIKU_MODEL }),
+      computeCorpusGeneration({
+        crMode: resolution.mode,
+        haikuModel: args.haikuModel ?? DEFAULT_HAIKU_MODEL,
+        synopsisDocMaxChars: resolution.mode === 'per_chunk_synopsis' ? SYNOPSIS_DOC_MAX_CHARS : undefined,
+      }),
     );
     return { kind: 'skipped', reason: 'no_chunks' };
   }
@@ -282,6 +298,7 @@ export async function reembedPageWithContextualRetrieval(
       const corpus_generation = computeCorpusGeneration({
         crMode: attemptMode,
         haikuModel,
+        synopsisDocMaxChars: attemptMode === 'per_chunk_synopsis' ? SYNOPSIS_DOC_MAX_CHARS : undefined,
       });
 
       // ── PHASE 2: single DB transaction ───────────────────────────
diff --git a/src/core/import-file.ts b/src/core/import-file.ts
index 4e7722abb..f988ee1cf 100644
--- a/src/core/import-file.ts
+++ b/src/core/import-file.ts
@@ -733,6 +733,11 @@ export async function importFromContent(
       : computeCorpusGeneration({
           crMode: effectiveCRMode,
           haikuModel: 'anthropic:claude-haiku-4-5-20251001',
+          // Inline import-file path never uses per_chunk_synopsis (refuses
+          // upstream); pass undefined so the doc-cap field stays out of
+          // the hash here. Per_chunk_synopsis runs through the Minion
+          // backfill handler which threads SYNOPSIS_DOC_MAX_CHARS through
+          // the service layer.
         });
 
   // Transaction wraps all DB writes. Every per-page tx call carries the
diff --git a/src/core/page-summary.ts b/src/core/page-summary.ts
index c4d7deedc..78e04d41c 100644
--- a/src/core/page-summary.ts
+++ b/src/core/page-summary.ts
@@ -44,6 +44,33 @@ const HAIKU_MAX_TOKENS = 200;
 /** Default model when caller doesn't override. Resolves through the gateway. */
 const DEFAULT_SYNOPSIS_MODEL = 'anthropic:claude-haiku-4-5-20251001';
 
+/**
+ * Hard cap on `documentText` length (chars) before send.
+ *
+ * 2026-05-25 fix wave: small local chat models (Gemma 4 E2B, Qwen3 4B) get
+ * dramatically slower on long contexts even with 131K-token windows declared.
+ * A 73K-char page synopsis on Gemma 4 E2B takes 60-120s, exceeding the
+ * worker's default 30s `lockDuration` and tripping `lock-lost` errors.
+ *
+ * Truncate to a budget that fits a small model's effective throughput while
+ * preserving enough document context for the synopsis to be useful. Truncates
+ * the TAIL because the head (title, frontmatter, intro) carries the
+ * document-level anchor the synopsis needs.
+ *
+ * Override per workload via `GBRAIN_SYNOPSIS_DOC_MAX_CHARS`. Default 32768
+ * (~8K tokens at 4 chars/tok) keeps small-model synopsis under ~30s.
+ * Anthropic Haiku is unaffected at this cap; bump higher when running
+ * frontier models if you want richer document anchoring.
+ */
+export const SYNOPSIS_DOC_MAX_CHARS = (() => {
+  const env = process.env.GBRAIN_SYNOPSIS_DOC_MAX_CHARS;
+  if (env && /^\d+$/.test(env)) {
+    const n = parseInt(env, 10);
+    if (n >= 512 && n <= 1_048_576) return n;
+  }
+  return 32768;
+})();
+
 /**
  * Synopsis prompt version. Folded into corpus_generation so prompt edits
  * invalidate prior embeddings via the v0.40.3.0 query_cache.page_generations
@@ -188,11 +215,19 @@ function buildUserPrompt(
   documentText: string,
   chunkText: string,
 ): string {
+  // Tail-truncate `documentText` to `SYNOPSIS_DOC_MAX_CHARS` so small local
+  // chat models don't stall on >100KB pages. Head preserved (title block,
+  // frontmatter, intro paragraphs carry the document-level anchor).
+  let trimmedDoc = documentText;
+  if (documentText.length > SYNOPSIS_DOC_MAX_CHARS) {
+    trimmedDoc = documentText.slice(0, SYNOPSIS_DOC_MAX_CHARS) +
+      `\n\n[... ${documentText.length - SYNOPSIS_DOC_MAX_CHARS} chars truncated for synopsis budget ...]`;
+  }
   return [
     `<page_title>${pageTitle}</page_title>`,
     '',
     '<full_document>',
-    documentText,
+    trimmedDoc,
     '</full_document>',
     '',
     '<chunk>',