fix(embed): thread write column into sumStaleChunkChars — sync cost gate followed the legacy predicate (#1262)

Review follow-up: the PR threaded the write-side column through
countStaleChunks/listStaleChunks but not sumStaleChunkChars, so the
sync cost gate counted an alt-column brain's fully-embedded corpus as
phantom backlog on every gate (inflating the --full estimate and the
deferred-mode backlog note). Both engines already share
buildStaleChunkWhere, so this is a type widening + one call-site
thread + a contrast assertion.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

src/commands/embed.ts

@@ -1,6 +1,7 @@
 import type { BrainEngine } from '../core/engine.ts';
 import { embedBatch, currentEmbeddingSignature } from '../core/embedding.ts';
-import type { ChunkInput } from '../core/types.ts';
+import type { ChunkInput, ResolvedColumn } from '../core/types.ts';
+import { resolveWriteColumnForEngine } from '../core/search/embedding-column.ts';
 import { chunkText } from '../core/chunkers/recursive.ts';
 import { createProgress, type ProgressReporter } from '../core/progress.ts';
 import { getCliOptions, cliOptsToProgressOptions } from '../core/cli-options.ts';
@@ -183,8 +184,13 @@ export class EmbeddingDimMismatchError extends Error {
  * fresh-install bug class at the very first invocation instead of letting
  * the worker pool hammer N pages with raw 22000 errors.
  */
-async function preflightDimMismatch(engine: BrainEngine, dryRun: boolean): Promise<void> {
+async function preflightDimMismatch(engine: BrainEngine, dryRun: boolean, embeddingColumn?: ResolvedColumn): Promise<void> {
   if (dryRun) return; // dry-run never embeds, no risk
+  // #1262: an alt-column brain writes to `embeddingColumn`, not the legacy
+  // `embedding` column — the legacy column's dims are irrelevant, and the
+  // registry entry (validated at resolve time) pins the target's dims. Only
+  // the legacy default path needs the schema-vs-gateway dim comparison.
+  if (embeddingColumn && embeddingColumn.name !== 'embedding') return;
   const { readContentChunksEmbeddingDim, embeddingMismatchMessage } = await import('../core/embedding-dim-check.ts');
   const { getEmbeddingDimensions, getEmbeddingModel } = await import('../core/ai/gateway.ts');
   let existing;
@@ -238,7 +244,12 @@ export async function runEmbedCore(engine: BrainEngine, opts: EmbedOpts): Promis
   // v0.37.11.0 (Lane D.2): pre-flight dim-mismatch check. Catches the headline
   // fresh-install bug class before the worker pool spends 20 parallel calls
   // hitting raw Postgres dimension errors.
-  await preflightDimMismatch(engine, !!opts.dryRun);
+  // #1262: resolve the write-side embedding column ONCE at the boundary
+  // (merged config + gateway model) and thread the descriptor through every
+  // upsertChunks / stale-scan below. undefined => legacy `embedding` column.
+  const embeddingColumn = await resolveWriteColumnForEngine(engine);
+
+  await preflightDimMismatch(engine, !!opts.dryRun, embeddingColumn);
 
   const result: EmbedResult = {
     embedded: 0,
@@ -253,7 +264,7 @@ export async function runEmbedCore(engine: BrainEngine, opts: EmbedOpts): Promis
     for (const s of opts.slugs) {
       if (isAborted(opts.signal)) break; // #1737: stop the per-slug loop on abort
       try {
-        await embedPage(engine, s, !!opts.dryRun, result, opts.sourceId, opts.signal);
+        await embedPage(engine, s, !!opts.dryRun, result, opts.sourceId, opts.signal, embeddingColumn);
       } catch (e: unknown) {
         serr(`  Error embedding ${s}: ${e instanceof Error ? e.message : e}`);
       }
@@ -347,7 +358,7 @@ export async function runEmbedCore(engine: BrainEngine, opts: EmbedOpts): Promis
         catchUp: opts.catchUp,
         pacer,
         paceMaxConcurrency,
-      }, opts.signal);
+      }, opts.signal, embeddingColumn);
     } finally {
       // E1: surface pacing telemetry (human + structured) when pacing was on.
       const snap = pacer.snapshot();
@@ -376,7 +387,7 @@ export async function runEmbedCore(engine: BrainEngine, opts: EmbedOpts): Promis
     return result;
   }
   if (opts.slug) {
-    await embedPage(engine, opts.slug, !!opts.dryRun, result, opts.sourceId, opts.signal);
+    await embedPage(engine, opts.slug, !!opts.dryRun, result, opts.sourceId, opts.signal, embeddingColumn);
     return result;
   }
   throw new Error('No embed target specified. Pass { slug }, { slugs }, { all }, or { stale }.');
@@ -521,8 +532,13 @@ async function embedPage(
   result: EmbedResult,
   sourceId?: string,
   signal?: AbortSignal,
+  embeddingColumn?: ResolvedColumn,
 ) {
   const opts = sourceId ? { sourceId } : undefined;
+  // #1262: write-side descriptor rides only on WRITE calls (upsertChunks).
+  const chunkOpts = (sourceId || embeddingColumn)
+    ? { ...(sourceId && { sourceId }), ...(embeddingColumn && { embeddingColumn }) }
+    : undefined;
   const page = await engine.getPage(slug, opts);
   if (!page) {
     throw new Error(`Page not found: ${slug}`);
@@ -554,7 +570,7 @@ async function embedPage(
     }
 
     if (inputs.length > 0) {
-      await engine.upsertChunks(slug, inputs, opts);
+      await engine.upsertChunks(slug, inputs, chunkOpts);
       chunks = await engine.getChunks(slug, opts);
     }
   }
@@ -589,7 +605,7 @@ async function embedPage(
     token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
   }));
 
-  await engine.upsertChunks(slug, updated, opts);
+  await engine.upsertChunks(slug, updated, chunkOpts);
   // v0.41.31: stamp provenance so a later model/dims swap is detectable as
   // stale. embedPage is the per-slug path used by `gbrain embed <slug>` AND
   // by `gbrain sync`'s post-import embed step (runEmbedCore({slugs})).
@@ -622,6 +638,7 @@ async function embedAll(
     paceMaxConcurrency?: number;
   },
   signal?: AbortSignal,
+  embeddingColumn?: ResolvedColumn,
 ) {
   // v0.41.31: current embedding provenance signature. Stamped onto pages
   // when their chunks are (re)embedded so a later model/dimension swap is
@@ -644,7 +661,7 @@ async function embedAll(
     // D7: thread sourceId so `gbrain embed --stale --source X` actually scopes.
     // v0.41.18.0 (A13): thread batchSize/priority/catchUp into the stale path.
     // #1737: thread the external abort signal so the cycle embed phase bails.
-    return await embedAllStale(engine, sourceId, dryRun, result, onProgress, staleOpts, signature, signal);
+    return await embedAllStale(engine, sourceId, dryRun, result, onProgress, staleOpts, signature, signal, embeddingColumn);
   }
 
   // --all path: pacer (no-op when off). E-1: lower the worker count to the
@@ -725,7 +742,10 @@ async function embedAll(
         embedding: embeddingMap.get(c.chunk_index) ?? undefined,
         token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
       }));
-      await observed(pacer, () => engine.upsertChunks(page.slug, updated, pageOpts));
+      await observed(pacer, () => engine.upsertChunks(page.slug, updated, {
+        ...(pageSourceId && { sourceId: pageSourceId }),
+        ...(embeddingColumn && { embeddingColumn }),
+      }));
       // v0.41.31: stamp embedding provenance so a later model swap is
       // detectable as stale.
       await observed(pacer, () =>
@@ -805,10 +825,16 @@ async function embedAllStale(
   },
   signature?: string,
   externalSignal?: AbortSignal,
+  embeddingColumn?: ResolvedColumn,
 ) {
   // D7: thread sourceId so source-scoped runs only count + visit
   // that source's NULL embeddings.
-  const sourceOpt = sourceId ? { sourceId } : undefined;
+  // #1262: the stale predicate follows the write-side column — without it an
+  // alt-column brain would perpetually re-select (and re-pay for) chunks whose
+  // target column is already populated.
+  const sourceOpt = (sourceId || embeddingColumn)
+    ? { ...(sourceId && { sourceId }), ...(embeddingColumn && { embeddingColumn }) }
+    : undefined;
 
   // v0.41.31: re-embed pages whose embedding_signature drifted (model/dims
   // swap). dry-run must NOT mutate, so it counts signature-stale via the
@@ -967,6 +993,7 @@ async function embedAllStale(
             afterUpdatedAt,
           }),
           ...(sourceId && { sourceId }),
+          ...(embeddingColumn && { embeddingColumn }),
         }),
       );
       if (batch.length === 0) {
@@ -1019,7 +1046,10 @@ async function embedAllStale(
             embedding: staleIdxToEmbedding.get(c.chunk_index) ?? undefined,
             token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
           }));
-          await observed(pacer, () => engine.upsertChunks(slug, merged, { sourceId: keySourceId }));
+          await observed(pacer, () => engine.upsertChunks(slug, merged, {
+            sourceId: keySourceId,
+            ...(embeddingColumn && { embeddingColumn }),
+          }));
           // v0.41.31: stamp provenance after the page's chunks are embedded —
           // but only when EVERY chunk was stale (fully re-embedded this pass).
           // A partially-stale page keeps preserved chunks of unknown/old
@@ -1090,7 +1120,7 @@ async function embedAllStale(
   // as a clean run — re-running won't help until the underlying failure is fixed.
   if (staleOpts?.catchUp && !effectiveSignal.aborted && embedFailures > 0) {
     const remaining = await engine.countStaleChunks(
-      signature ? { signature, ...(sourceId ? { sourceId } : {}) } : (sourceId ? { sourceId } : undefined),
+      signature ? { signature, ...sourceOpt } : sourceOpt,
     );
     if (remaining > 0) {
       serr(`\n  [embed] catch-up finished but ${remaining} chunk(s) remain stale after ${embedFailures} embed failure(s). These are not embeddable as-is; re-running won't clear them until the underlying error is resolved.`);

src/commands/sync.ts

@@ -576,9 +576,16 @@ async function runInlineCostGate(
 
   // Stale backlog: cheap single SQL; fail-open to 0 so a transient DB hiccup
   // never blocks the sync. Signature-aware (model/dims swap surfaces here).
+  // #1262: follow the write-side embedding column — otherwise an alt-column
+  // brain's fully-embedded corpus counts as phantom backlog on every gate.
   let staleChars = 0;
   try {
-    staleChars = await engine.sumStaleChunkChars({ signature: currentEmbeddingSignature() });
+    const { resolveWriteColumnForEngine } = await import('../core/search/embedding-column.ts');
+    const embeddingColumn = await resolveWriteColumnForEngine(engine);
+    staleChars = await engine.sumStaleChunkChars({
+      signature: currentEmbeddingSignature(),
+      ...(embeddingColumn && { embeddingColumn }),
+    });
   } catch {
     staleChars = 0;
   }

src/core/chunkers/code.ts

@@ -18,9 +18,8 @@
  * at runtime.
  */
 
-import { chunkText as recursiveChunk, capByEstimatedTokens, DEFAULT_MAX_EST_TOKENS } from './recursive.ts';
+import { chunkText as recursiveChunk } from './recursive.ts';
 import { buildQualifiedName } from './qualified-names.ts';
-import { estimateEmbeddingTokens } from '../cjk.ts';
 
 // Embed the tree-sitter runtime + per-language grammars as files.
 // `with { type: 'file' }` returns a path (string) at runtime. Bun bundles
@@ -112,15 +111,7 @@ import G_ZIG from '../../assets/wasm/grammars/tree-sitter-zig.wasm' with { type:
 // chunks get the new columns populated. Without this, the v28 backfill
 // gives every existing chunk a search_vector but subsequent Layer 5 AST
 // work would silently no-op.
-//
-// v5: estimated-token hard cap on AST-path chunks (capCodeChunks). A node
-// splitLargeNode can't subdivide (giant single-statement function, huge
-// literal) previously shipped WHOLE regardless of size and could overflow
-// strict per-request embedding-token limits (local llama-server crashes
-// past ~2,050 tokens, measured). Mirrors the markdown
-// chunker's v4 cap; fallback-path chunks are already capped inside
-// recursiveChunk.
-export const CHUNKER_VERSION = 5;
+export const CHUNKER_VERSION = 4;
 
 // Lazy-loaded tree-sitter module (v0.22.x API: Parser is default export)
 let Parser: typeof import('web-tree-sitter') | null = null;
@@ -717,7 +708,7 @@ export async function chunkCodeTextFull(
     if (chunks.length === 0) {
       return { chunks: fallbackChunks(source, filePath, language, opts), edges: rawEdges };
     }
-    return { chunks: capCodeChunks(mergeSmallSiblings(chunks, chunkTarget)), edges: rawEdges };
+    return { chunks: mergeSmallSiblings(chunks, chunkTarget), edges: rawEdges };
   } catch {
     return { chunks: fallbackChunks(source, filePath, language, opts), edges: [] };
   } finally {
@@ -800,33 +791,6 @@ function mergeSmallSiblings(chunks: CodeChunk[], chunkTarget: number): CodeChunk
   return merged;
 }
 
-/**
- * v5 final safety pass for AST-path chunks: split any chunk whose
- * ESTIMATED embedding tokens (conservative per-char-class heuristic,
- * cjk.ts) exceed DEFAULT_MAX_EST_TOKENS. Reaches chunks the AST logic
- * can't subdivide — splitLargeNode returns [] for nodes with < 2 body
- * children (giant single-statement functions, huge literals), which
- * previously shipped whole at any size.
- *
- * Split pieces inherit the source chunk's metadata verbatim; start/end
- * lines become approximate for pieces after the first. Acceptable —
- * these chunks exist for embedding + retrieval, and the alternative was
- * an embedding request the server rejects (or worse, crashes on).
- */
-function capCodeChunks(chunks: CodeChunk[]): CodeChunk[] {
-  if (chunks.every((c) => estimateEmbeddingTokens(c.text) <= DEFAULT_MAX_EST_TOKENS)) {
-    return chunks;
-  }
-  const out: CodeChunk[] = [];
-  for (const c of chunks) {
-    const pieces = capByEstimatedTokens(c.text, DEFAULT_MAX_EST_TOKENS);
-    for (const piece of pieces) {
-      out.push({ ...c, text: piece, index: out.length, metadata: { ...c.metadata } });
-    }
-  }
-  return out;
-}
-
 function buildMergedChunk(group: CodeChunk[], index: number): CodeChunk {
   const first = group[0]!;
   const last = group[group.length - 1]!;

src/core/chunkers/recursive.ts

@@ -17,13 +17,7 @@
  * Lossless invariant: non-overlapping portions reassemble to original.
  */
 
-import {
-  countCJKAwareWords,
-  CJK_SENTENCE_DELIMITERS,
-  CJK_CLAUSE_DELIMITERS,
-  charEmbedTokenWeight,
-  estimateEmbeddingTokens,
-} from '../cjk.ts';
+import { countCJKAwareWords, CJK_SENTENCE_DELIMITERS, CJK_CLAUSE_DELIMITERS } from '../cjk.ts';
 
 /**
  * Markdown chunker version. Folded into the per-page chunker_version column
@@ -39,20 +33,8 @@ import {
  * re-embed (not re-chunk) so existing pages pick up the wrapper on the
  * post-upgrade reembed sweep. See
  * `src/core/contextual-retrieval-service.ts`.
- *
- * v4: estimated-token hard cap + whitespace-word undercount fix. The word
- * pipeline counted a 150-char URL as ONE whitespace word, so URL/phone/
- * email-dense docs (CJK density < 0.30 → whitespace fallback) produced
- * 3-4K-char chunks that overflow strict per-request embedding-token
- * limits (measured: local llama-server crashes past ~2,050 tokens; URL
- * soup tokenizes at ~1.6 chars/token). Two changes:
- *   1. countWords() floors the count at ceil(nonWhitespaceChars/6) so a
- *      URL counts roughly per-character, not as one word.
- *   2. capByEstimatedTokens() final pass guarantees every chunk fits
- *      `maxTokens` (default 1500) under a conservative per-char-class
- *      token estimate, regardless of how word counting misjudged it.
  */
-export const MARKDOWN_CHUNKER_VERSION = 4;
+export const MARKDOWN_CHUNKER_VERSION = 3;
 
 const DELIMITERS: string[][] = [
   ['\n\n'],                          // L0: paragraphs
@@ -66,20 +48,8 @@ export interface ChunkOptions {
   chunkSize?: number;    // target words per chunk (default 300)
   chunkOverlap?: number; // overlap words (default 50)
   maxChars?: number;     // hard cap on any chunk's char length (default 6000)
-  /**
-   * v4: hard cap on any chunk's ESTIMATED embedding tokens (default 1500).
-   * Estimate = conservative per-char-class weights (see cjk.ts
-   * estimateEmbeddingTokens) — deliberately high, so the real tokenizer
-   * count stays below this value. Default leaves headroom for the
-   * contextual-retrieval wrapper (≤ ~630 chars) under a ~2,050-token
-   * per-request embedding server limit.
-   */
-  maxTokens?: number;
 }
 
-/** v4 default for ChunkOptions.maxTokens — see the field doc above. */
-export const DEFAULT_MAX_EST_TOKENS = 1500;
-
 export interface TextChunk {
   text: string;
   index: number;
@@ -103,7 +73,6 @@ export function chunkText(text: string, opts?: ChunkOptions): TextChunk[] {
   const chunkSize = opts?.chunkSize || 300;
   const chunkOverlap = opts?.chunkOverlap || 50;
   const maxChars = opts?.maxChars || 6000;
-  const maxTokens = opts?.maxTokens || DEFAULT_MAX_EST_TOKENS;
 
   if (!text || text.trim().length === 0) return [];
 
@@ -120,9 +89,8 @@ export function chunkText(text: string, opts?: ChunkOptions): TextChunk[] {
 
   const wordCount = countWords(stripped);
   if (wordCount <= chunkSize) {
-    // Single-chunk path: still apply the maxChars + maxTokens caps.
-    const capped = capByChars(stripped.trim(), maxChars)
-      .flatMap((t) => capByEstimatedTokens(t, maxTokens));
+    // Single-chunk path: still apply the maxChars cap.
+    const capped = capByChars(stripped.trim(), maxChars);
     return capped.map((t, i) => ({ text: t, index: i }));
   }
 
@@ -133,14 +101,9 @@ export function chunkText(text: string, opts?: ChunkOptions): TextChunk[] {
   // v0.32.7: hard char cap. Catches pathological CJK + whitespace-less text
   // that the word-level pipeline can't bound (a single Chinese paragraph can
   // exceed 8192 OpenAI embedding tokens at any word count).
-  // v4: estimated-token cap on top — the char cap alone passes token-dense
-  // content (URL soup at ~1.6 chars/token) that overflows strict embedding
-  // server limits.
   const capped: string[] = [];
   for (const chunk of withOverlap) {
-    for (const piece of capByChars(chunk.trim(), maxChars)) {
-      capped.push(...capByEstimatedTokens(piece, maxTokens));
-    }
+    capped.push(...capByChars(chunk.trim(), maxChars));
   }
   return capped.map((t, i) => ({ text: t, index: i }));
 }
@@ -169,68 +132,6 @@ function capByChars(text: string, maxChars: number): string[] {
   return out;
 }
 
-/**
- * How far back (in chars) the token cap looks for a friendly cut point
- * before falling back to a hard cut. 300 covers typical rollup/list line
- * lengths so forced splits land at line starts, not mid-URL.
- */
-const TOKEN_CAP_CUT_LOOKBACK = 300;
-
-/**
- * v4: hard-cap a chunk's ESTIMATED embedding tokens. Final safety pass —
- * runs after capByChars on every chunk, so no upstream miscounting
- * (whitespace-word fallback, overlap inflation, char-cap survivors) can
- * emit a chunk past `maxTokens`.
- *
- * Cut placement prefers, within the last TOKEN_CAP_CUT_LOOKBACK chars of
- * the window: a newline, then any whitespace, then a hard cut. This keeps
- * forced splits off mid-line/mid-URL positions for list-shaped content
- * and inside code fences. No overlap is added (pieces stay lossless
- * modulo the trims the char cap already applies).
- *
- * @internal exported for the code chunker (code.ts) and tests.
- */
-export function capByEstimatedTokens(text: string, maxTokens: number): string[] {
-  if (text.length === 0) return [];
-  if (estimateEmbeddingTokens(text) <= maxTokens) return [text];
-
-  const out: string[] = [];
-  let start = 0;
-  while (start < text.length) {
-    // Greedily extend the window until the next char would break the cap.
-    // Always take at least one char so the loop makes forward progress.
-    let est = 0;
-    let end = start;
-    while (end < text.length) {
-      const w = charEmbedTokenWeight(text.charCodeAt(end));
-      if (est + w > maxTokens && end > start) break;
-      est += w;
-      end++;
-    }
-
-    if (end < text.length) {
-      const windowStart = Math.max(start + 1, end - TOKEN_CAP_CUT_LOOKBACK);
-      let cut = text.lastIndexOf('\n', end - 1);
-      if (cut < windowStart) {
-        cut = -1;
-        for (let i = end - 1; i >= windowStart; i--) {
-          const code = text.charCodeAt(i);
-          if (code === 0x20 || (code >= 0x09 && code <= 0x0d)) {
-            cut = i;
-            break;
-          }
-        }
-      }
-      if (cut >= windowStart) end = cut + 1;
-    }
-
-    const slice = text.slice(start, end).trim();
-    if (slice.length > 0) out.push(slice);
-    start = end;
-  }
-  return out;
-}
-
 function recursiveSplit(text: string, level: number, target: number): string[] {
   if (level >= DELIMITERS.length) {
     // Level 4: split on whitespace
@@ -416,19 +317,7 @@ function extractTrailingContext(text: string, targetWords: number): string {
  * Delegated to src/core/cjk.ts so the slugify whitelist, expansion
  * detection, and PGLite keyword fallback all agree on what "CJK enough"
  * means.
- *
- * v4: floored at ceil(nonWhitespaceChars/6). The whitespace fallback
- * counts a 150-char URL as ONE word, so URL/phone/email-dense docs
- * (whose ASCII mass pushes CJK density below the 0.30 threshold) were
- * sized at a fraction of their real bulk and merged into 3-4K-char
- * chunks. The floor makes long whitespace-less runs count roughly
- * per-character while leaving normal Latin prose untouched (average
- * English word ≈ 5 chars < 6, so the whitespace count still wins).
- * Kept local to the chunker — search/expansion.ts keeps the original
- * countCJKAwareWords semantics for its query-length check.
  */
 function countWords(text: string): number {
-  const cjkAware = countCJKAwareWords(text);
-  const nonWhitespace = text.replace(/\s/g, '').length;
-  return Math.max(cjkAware, Math.ceil(nonWhitespace / 6));
+  return countCJKAwareWords(text);
 }

src/core/cjk.ts

@@ -65,64 +65,3 @@ export function countCJKAwareWords(s: string): number {
 export function escapeLikePattern(s: string): string {
   return s.replace(/\\/g, '\\\\').replace(/%/g, '\\%').replace(/_/g, '\\_');
 }
-
-/**
- * Conservative per-char-class embedding-token weights (markdown chunker v4).
- *
- * Why this exists: the chunker's "word" counting drastically UNDER-counts
- * whitespace-less ASCII runs (a 150-char URL = 1 whitespace word), so
- * word-based size targets can emit chunks that overflow an embedding
- * server's per-request token limit. Measured on a local Qwen3-embedding
- * llama-server stack:
- *   - URL/phone/email-dense text tokenizes at ~1.6 chars/token
- *   - base64-ish / minified blobs approach ~1.3 chars/token (worst case)
- *   - Korean prose tokenizes NO WORSE than 1 char/token in practice
- *
- * Weights are deliberately HIGH (tokens are overestimated) so any cap
- * based on this estimate is safe against real tokenizers:
- *   - CJK char        → 1.0 token  (real CJK prose is cheaper)
- *   - other non-space → 0.75 token (≈1.33 chars/token, covers base64)
- *   - whitespace      → 0.1 token  (mostly folds into neighbor tokens)
- */
-export const EMBED_TOKEN_WEIGHT_CJK = 1.0;
-export const EMBED_TOKEN_WEIGHT_OTHER = 0.75;
-export const EMBED_TOKEN_WEIGHT_WS = 0.1;
-
-/** BMP CJK check by UTF-16 code unit — same ranges as CJK_SLUG_CHARS. */
-export function isCJKCodeUnit(code: number): boolean {
-  return (
-    (code >= 0x4e00 && code <= 0x9fff) || // Han
-    (code >= 0x3040 && code <= 0x309f) || // Hiragana
-    (code >= 0x30a0 && code <= 0x30ff) || // Katakana
-    (code >= 0xac00 && code <= 0xd7af)    // Hangul Syllables
-  );
-}
-
-/**
- * Per-code-unit token weight. Unrecognized whitespace (exotic Unicode
- * spaces) intentionally falls into OTHER — that only overestimates.
- */
-export function charEmbedTokenWeight(code: number): number {
-  if (isCJKCodeUnit(code)) return EMBED_TOKEN_WEIGHT_CJK;
-  if (
-    code === 0x20 || (code >= 0x09 && code <= 0x0d) ||
-    code === 0xa0 || code === 0x3000
-  ) {
-    return EMBED_TOKEN_WEIGHT_WS;
-  }
-  return EMBED_TOKEN_WEIGHT_OTHER;
-}
-
-/**
- * Tokenizer-free embedding-token estimate (conservative overestimate).
- * See weight docs above. Astral chars count as 2 OTHER code units —
- * another overestimate, which is the safe direction.
- */
-export function estimateEmbeddingTokens(s: string): number {
-  if (s.length === 0) return 0;
-  let est = 0;
-  for (let i = 0; i < s.length; i++) {
-    est += charEmbedTokenWeight(s.charCodeAt(i));
-  }
-  return Math.ceil(est);
-}

src/core/contextual-retrieval-service.ts

@@ -61,6 +61,7 @@ import {
   type SynopsisFailureKind,
 } from './audit-synopsis.ts';
 import type { BrainEngine } from './engine.ts';
+import { resolveWriteColumnForEngine } from './search/embedding-column.ts';
 import type { ChunkInput, CRMode, Page } from './types.ts';
 import type { SourceRow } from './sources-ops.ts';
 
@@ -286,9 +287,13 @@ export async function reembedPageWithContextualRetrieval(
 
       // ── PHASE 2: single DB transaction ───────────────────────────
       try {
+        // #1262: contextual re-embeds write TEXT embeddings — thread the
+        // caller-resolved write column like every other embed path.
+        const embeddingColumn = await resolveWriteColumnForEngine(args.engine);
         await args.engine.transaction(async (tx) => {
           await tx.upsertChunks(args.pageSlug, phase1.embeddedChunks, {
             sourceId: args.sourceId,
+            ...(embeddingColumn && { embeddingColumn }),
           });
           await tx.updatePageContextualRetrievalState(
             args.pageSlug,

src/core/embed-stale.ts

@@ -18,7 +18,7 @@
  */
 
 import type { BrainEngine } from './engine.ts';
-import type { ChunkInput } from './types.ts';
+import type { ChunkInput, ResolvedColumn } from './types.ts';
 import { embedBatchWithBackoff } from '../commands/embed.ts';
 import { type DbPacer, createNoopPacer, observed } from './db-pacer.ts';
 import { AbortError } from './abort-check.ts';
@@ -61,6 +61,13 @@ export interface EmbedStaleOpts {
    * Omit to keep the legacy `embedding IS NULL`-only behavior.
    */
   embeddingSignature?: string;
+  /**
+   * #1262: caller-resolved write-side embedding column. Threaded into BOTH
+   * listStaleChunks (staleness predicate) and upsertChunks (write target) so
+   * an alt-column brain converges instead of re-selecting embedded rows.
+   * Resolve at the boundary via `resolveWriteColumnForEngine()`.
+   */
+  embeddingColumn?: ResolvedColumn;
   /**
    * DB-contention pacer (paced-backfill). When enabled it (a) supplies the
    * worker count via the caller passing `concurrency = bundle.maxConcurrency`
@@ -156,6 +163,7 @@ export async function embedStaleForSource(
         afterPageId,
         afterChunkIndex,
         sourceId,
+        ...(opts.embeddingColumn && { embeddingColumn: opts.embeddingColumn }),
       }),
     );
     if (batch.length === 0) {
@@ -223,7 +231,10 @@ export async function embedStaleForSource(
           doc_comment: c.doc_comment ?? undefined,
           symbol_name_qualified: c.symbol_name_qualified ?? undefined,
         }));
-        await observed(pacer, () => engine.upsertChunks(slug, merged, { sourceId: keySourceId }));
+        await observed(pacer, () => engine.upsertChunks(slug, merged, {
+          sourceId: keySourceId,
+          ...(opts.embeddingColumn && { embeddingColumn: opts.embeddingColumn }),
+        }));
         // v0.41.31: stamp provenance only when EVERY chunk was stale (fully
         // re-embedded this pass) — a partially-stale page keeps preserved
         // chunks of unknown provenance, so don't claim current. After the

src/core/engine.ts

@@ -12,6 +12,7 @@ import type {
   BrainStats, BrainHealth,
   IngestLogEntry, IngestLogInput,
   EngineConfig,
+  ResolvedColumn,
   CodeEdgeInput, CodeEdgeResult,
   EvalCandidate, EvalCandidateInput,
   EvalCaptureFailure, EvalCaptureFailureReason,
@@ -987,8 +988,13 @@ export interface BrainEngine {
    * — Postgres rolls back automatically on conn drop, so commit-ambiguous
    * failure replays to the same end state. Callers MUST NOT wrap externally;
    * see {@link BatchOpts} retry-contract block.
+   *
+   * `opts.embeddingColumn` (optional) selects the content_chunks column that
+   * receives TEXT embeddings (#1262). The caller resolves the descriptor at
+   * the import/embed boundary via `resolveWriteColumn()`; engines never read
+   * config or choose columns themselves. Omitted => legacy `embedding`.
    */
-  upsertChunks(slug: string, chunks: ChunkInput[], opts?: { sourceId?: string } & BatchOpts): Promise<void>;
+  upsertChunks(slug: string, chunks: ChunkInput[], opts?: { sourceId?: string; embeddingColumn?: ResolvedColumn } & BatchOpts): Promise<void>;
   /**
    * Read every chunk for a page. `opts.sourceId` source-scopes the page
    * lookup; without it, multi-source brains return chunks from every
@@ -1005,8 +1011,13 @@ export interface BrainEngine {
    * counts across every source in the brain. Operators running
    * `gbrain embed --stale --source media-corpus` expect only that
    * source's NULLs touched; the caller threads `sourceId` here.
+   *
+   * `opts.embeddingColumn` switches the staleness predicate from the legacy
+   * `embedding` column to the resolved write-side column, so alt-column
+   * brains do not perpetually re-select rows whose target column is already
+   * populated (#1262). Must match the eventual upsertChunks target.
    */
-  countStaleChunks(opts?: { sourceId?: string; signature?: string }): Promise<number>;
+  countStaleChunks(opts?: { sourceId?: string; signature?: string; embeddingColumn?: ResolvedColumn }): Promise<number>;
   /**
    * Sum of LENGTH(chunk_text) over stale chunks — the character-count
    * backlog the embed phase / embed-backfill will process. Sibling of
@@ -1020,8 +1031,13 @@ export interface BrainEngine {
    * model signature (a model/dims swap). NULL signature is GRANDFATHERED
    * (never counted) so the post-migration corpus isn't flagged en masse.
    * Omit `signature` for the legacy `embedding IS NULL`-only count.
+   *
+   * `opts.embeddingColumn` switches the staleness predicate to the resolved
+   * write-side column (#1262) — same contract as countStaleChunks — so the
+   * sync cost gate doesn't count an alt-column brain's fully-embedded corpus
+   * as phantom backlog.
    */
-  sumStaleChunkChars(opts?: { sourceId?: string; signature?: string }): Promise<number>;
+  sumStaleChunkChars(opts?: { sourceId?: string; signature?: string; embeddingColumn?: ResolvedColumn }): Promise<number>;
   /**
    * Stamp `pages.embedding_signature = signature` for one page. Called after
    * a page's chunks are (re)embedded so a later model swap can detect it as
@@ -1069,6 +1085,9 @@ export interface BrainEngine {
     // both round-trip TIMESTAMPTZ as Date | string; ISO string is the
     // common denominator on the wire).
     afterUpdatedAt?: string | null;
+    // #1262: staleness predicate targets this column when set (must match
+    // countStaleChunks and the eventual upsertChunks write target).
+    embeddingColumn?: ResolvedColumn;
   }): Promise<StaleChunkRow[]>;
   /**
    * Delete every chunk for a page. Internal page-id lookup is sourceId-scoped

src/core/import-file.ts

@@ -10,7 +10,8 @@ import { findChunkForOffset } from './chunkers/edge-extractor.ts';
 import { extractCodeRefs, imageOfCandidates } from './link-extraction.ts';
 import { embedBatch, embedMultimodal, currentEmbeddingSignature } from './embedding.ts';
 import { slugifyPath, slugifyCodePath, isCodeFilePath } from './sync.ts';
-import type { ChunkInput, PageInput, PageType } from './types.ts';
+import type { ChunkInput, PageInput, PageType, ResolvedColumn } from './types.ts';
+import { resolveWriteColumnForEngine } from './search/embedding-column.ts';
 import { computeEffectiveDate } from './effective-date.ts';
 import { MARKDOWN_CHUNKER_VERSION } from './chunkers/recursive.ts';
 import { logSlugFallback } from './audit-slug-fallback.ts';
@@ -740,6 +741,14 @@ export async function importFromContent(
   // schema DEFAULT — required for multi-source brains; harmless ('default')
   // for single-source callers.
   const txOpts = sourceId ? { sourceId } : undefined;
+  // #1262: resolve the write-side embedding column once (merged config +
+  // gateway model) BEFORE the transaction; the descriptor rides only on
+  // upsertChunks so text embeddings land in the registered column.
+  const chunkWriteColumn = await resolveWriteColumnForEngine(engine);
+  const chunkOpts: { sourceId?: string; embeddingColumn?: ResolvedColumn } | undefined =
+    (sourceId || chunkWriteColumn)
+      ? { ...(sourceId && { sourceId }), ...(chunkWriteColumn && { embeddingColumn: chunkWriteColumn }) }
+      : undefined;
   await engine.transaction(async (tx) => {
     if (existing) await tx.createVersion(slug, txOpts);
 
@@ -824,7 +833,7 @@ export async function importFromContent(
     }
 
     if (chunks.length > 0) {
-      await tx.upsertChunks(slug, chunks, txOpts);
+      await tx.upsertChunks(slug, chunks, chunkOpts);
       // v0.41.31: stamp embedding provenance when this import actually
       // embedded (not --no-embed), so a later model/dims swap is detectable
       // as stale via embed --stale. The deferred/backfill + per-slug embed
@@ -1064,6 +1073,12 @@ export async function importCodeFile(
   const title = `${relativePath} (${lang})`;
   const sourceId = opts.sourceId;
   const txOpts = sourceId ? { sourceId } : undefined;
+  // #1262: write-side embedding column descriptor (rides only on upsertChunks).
+  const chunkWriteColumn = await resolveWriteColumnForEngine(engine);
+  const chunkOpts: { sourceId?: string; embeddingColumn?: ResolvedColumn } | undefined =
+    (sourceId || chunkWriteColumn)
+      ? { ...(sourceId && { sourceId }), ...(chunkWriteColumn && { embeddingColumn: chunkWriteColumn }) }
+      : undefined;
 
   const byteLength = Buffer.byteLength(content, 'utf-8');
   if (byteLength > MAX_FILE_SIZE) {
@@ -1183,7 +1198,7 @@ export async function importCodeFile(
     await tx.addTag(slug, lang, txOpts);
 
     if (chunks.length > 0) {
-      await tx.upsertChunks(slug, chunks, txOpts);
+      await tx.upsertChunks(slug, chunks, chunkOpts);
       // v0.41.31: stamp embedding provenance ONLY when every chunk was
       // freshly embedded with the current model this call (no reuse-by-hash
       // carrying old-model vectors). Mixed pages stay unstamped rather than
@@ -1332,6 +1347,12 @@ export async function withImportTransaction(
 ): Promise<void> {
   const sourceId = spec.sourceId ?? 'default';
   const txOpts = spec.sourceId ? { sourceId: spec.sourceId } : undefined;
+  // #1262: write-side embedding column descriptor (rides only on upsertChunks).
+  const chunkWriteColumn = await resolveWriteColumnForEngine(engine);
+  const chunkOpts: { sourceId?: string; embeddingColumn?: ResolvedColumn } | undefined =
+    (spec.sourceId || chunkWriteColumn)
+      ? { ...(spec.sourceId && { sourceId: spec.sourceId }), ...(chunkWriteColumn && { embeddingColumn: chunkWriteColumn }) }
+      : undefined;
   await engine.transaction(async (tx) => {
     if (spec.hadExisting) await tx.createVersion(spec.slug, txOpts);
     await tx.putPage(spec.slug, spec.page, txOpts);
@@ -1347,7 +1368,7 @@ export async function withImportTransaction(
     }
     if (spec.chunks !== undefined) {
       if (spec.chunks.length > 0) {
-        await tx.upsertChunks(spec.slug, spec.chunks, txOpts);
+        await tx.upsertChunks(spec.slug, spec.chunks, chunkOpts);
       } else {
         await tx.deleteChunks(spec.slug, txOpts);
       }

src/core/minions/handlers/embed-backfill.ts

@@ -35,6 +35,7 @@ import { tryAcquireDbLock } from '../../db-lock.ts';
 import { BudgetTracker, BudgetExhausted } from '../../budget/budget-tracker.ts';
 import { withBudgetTracker } from '../../ai/gateway.ts';
 import { embedStaleForSource } from '../../embed-stale.ts';
+import { resolveWriteColumnForEngine } from '../../search/embedding-column.ts';
 import { currentEmbeddingSignature } from '../../embedding.ts';
 import { type DbPacer, createDbPacer, createNoopPacer } from '../../db-pacer.ts';
 import { resolvePaceMode, loadPaceModeConfig, readPaceEnv } from '../../pace-mode.ts';
@@ -164,12 +165,16 @@ export function makeEmbedBackfillHandler(engine: BrainEngine) {
     // the supervisor, so pacing it is the headline win.
     const { pacer, concurrency } = await resolveBackfillPacer(engine, job.data);
 
+    // #1262: resolve the write-side embedding column once at the job boundary.
+    const embeddingColumn = await resolveWriteColumnForEngine(engine);
+
     try {
       const result = await withBudgetTracker(tracker, async () =>
         embedStaleForSource(engine, sourceId, {
           batchSize,
           signal: job.signal,
           pacer,
+          ...(embeddingColumn && { embeddingColumn }),
           ...(concurrency !== undefined && { concurrency }),
           // v0.41.31: re-embed pages whose model signature drifted + stamp
           // provenance as chunks land.

src/core/pglite-engine.ts

@@ -40,6 +40,7 @@ import type {
   BrainStats, BrainHealth,
   IngestLogEntry, IngestLogInput,
   EngineConfig,
+  ResolvedColumn,
   EvalCandidate, EvalCandidateInput,
   EvalCaptureFailure, EvalCaptureFailureReason,
   SalienceOpts, SalienceResult, AnomaliesOpts, AnomalyResult,
@@ -2230,12 +2231,20 @@ export class PGLiteEngine implements BrainEngine {
   }
 
   // Chunks
-  async upsertChunks(slug: string, chunks: ChunkInput[], opts?: { sourceId?: string } & BatchOpts): Promise<void> {
+  async upsertChunks(slug: string, chunks: ChunkInput[], opts?: { sourceId?: string; embeddingColumn?: ResolvedColumn } & BatchOpts): Promise<void> {
     return this.batchRetry(opts?.auditSite ?? 'upsertChunks', opts?.signal, () => this._upsertChunksOnce(slug, chunks, opts), chunks.length);
   }
 
-  private async _upsertChunksOnce(slug: string, chunks: ChunkInput[], opts?: { sourceId?: string }): Promise<void> {
+  private async _upsertChunksOnce(slug: string, chunks: ChunkInput[], opts?: { sourceId?: string; embeddingColumn?: ResolvedColumn }): Promise<void> {
     const sourceId = opts?.sourceId ?? 'default';
+    // #1262: caller-resolved write target for TEXT embeddings. Descriptor
+    // names are identifier-validated + quoted by buildVectorCastFragment;
+    // omitted => legacy `embedding vector`. Mirrors postgres-engine.ts.
+    const targetFragment = opts?.embeddingColumn
+      ? buildVectorCastFragment(opts.embeddingColumn)
+      : undefined;
+    const targetCol = targetFragment?.col ?? 'embedding';
+    const embeddingCast = targetFragment?.castSql.replace('$1::', '') ?? 'vector';
 
     // Source-scope the page-id lookup so duplicate slugs in different sources
     // do not return multiple rows or target the wrong page.
@@ -2270,7 +2279,7 @@ export class PGLiteEngine implements BrainEngine {
     // list. Image chunks pass embedding=null + embedding_image=Float32Array
     // (1024-dim Voyage). Text/code chunks pass embedding=Float32Array +
     // embedding_image=null. Default modality='text' when omitted.
-    const cols = '(page_id, chunk_index, chunk_text, chunk_source, embedding, model, token_count, embedded_at, language, symbol_name, symbol_type, start_line, end_line, parent_symbol_path, doc_comment, symbol_name_qualified, modality, embedding_image)';
+    const cols = `(page_id, chunk_index, chunk_text, chunk_source, ${targetCol}, model, token_count, embedded_at, language, symbol_name, symbol_type, start_line, end_line, parent_symbol_path, doc_comment, symbol_name_qualified, modality, embedding_image)`;
     const rowParts: string[] = [];
     const params: unknown[] = [];
     let paramIdx = 1;
@@ -2288,7 +2297,7 @@ export class PGLiteEngine implements BrainEngine {
       const modality = chunk.modality ?? 'text';
 
       // Inline ::vector NULL literals to avoid a per-branch placeholder.
-      const embeddingPh = embeddingStr ? `$${paramIdx++}::vector` : 'NULL';
+      const embeddingPh = embeddingStr ? `$${paramIdx++}::${embeddingCast}` : 'NULL';
       const embeddedAtPh = embeddingStr ? 'now()' : 'NULL';
       const embeddingImagePh = embeddingImageStr ? `$${paramIdx++}::vector` : 'NULL';
 
@@ -2327,19 +2336,19 @@ export class PGLiteEngine implements BrainEngine {
        ON CONFLICT (page_id, chunk_index) DO UPDATE SET
          chunk_text = EXCLUDED.chunk_text,
          chunk_source = EXCLUDED.chunk_source,
-         embedding = CASE
-           WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.embedding
-           WHEN content_chunks.embedding IS NULL THEN EXCLUDED.embedding
+         ${targetCol} = CASE
+           WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.${targetCol}
+           WHEN content_chunks.${targetCol} IS NULL THEN EXCLUDED.${targetCol}
            WHEN EXCLUDED.embedded_at IS NOT NULL
                 AND (content_chunks.embedded_at IS NULL OR EXCLUDED.embedded_at > content_chunks.embedded_at)
-                THEN EXCLUDED.embedding
-           ELSE content_chunks.embedding
+                THEN EXCLUDED.${targetCol}
+           ELSE content_chunks.${targetCol}
          END,
          model = COALESCE(EXCLUDED.model, content_chunks.model),
          token_count = EXCLUDED.token_count,
          embedded_at = CASE
-           WHEN EXCLUDED.chunk_text != content_chunks.chunk_text AND EXCLUDED.embedding IS NULL THEN NULL
-           WHEN content_chunks.embedding IS NULL AND EXCLUDED.embedding IS NOT NULL THEN EXCLUDED.embedded_at
+           WHEN EXCLUDED.chunk_text != content_chunks.chunk_text AND EXCLUDED.${targetCol} IS NULL THEN NULL
+           WHEN content_chunks.${targetCol} IS NULL AND EXCLUDED.${targetCol} IS NOT NULL THEN EXCLUDED.embedded_at
            WHEN EXCLUDED.embedded_at IS NOT NULL
                 AND (content_chunks.embedded_at IS NULL OR EXCLUDED.embedded_at > content_chunks.embedded_at)
                 THEN EXCLUDED.embedded_at
@@ -2377,14 +2386,19 @@ export class PGLiteEngine implements BrainEngine {
    * drift (NULL grandfathered → never stale). Shared by countStaleChunks +
    * sumStaleChunkChars so they can't drift.
    */
-  private buildStaleChunkWhere(opts?: { sourceId?: string; signature?: string }): { where: string; params: unknown[] } {
+  private buildStaleChunkWhere(opts?: { sourceId?: string; signature?: string; embeddingColumn?: ResolvedColumn }): { where: string; params: unknown[] } {
+    // #1262: staleness targets the caller-resolved write column when set
+    // (identifier-validated + quoted); legacy `embedding` otherwise.
+    const staleCol = opts?.embeddingColumn
+      ? buildVectorCastFragment(opts.embeddingColumn).col
+      : 'embedding';
     const params: unknown[] = [];
     const conds: string[] = [];
     if (opts?.signature !== undefined) {
       params.push(opts.signature);
-      conds.push(`(cc.embedding IS NULL OR (p.embedding_signature IS NOT NULL AND p.embedding_signature <> $${params.length}))`);
+      conds.push(`(cc.${staleCol} IS NULL OR (p.embedding_signature IS NOT NULL AND p.embedding_signature <> $${params.length}))`);
     } else {
-      conds.push(`cc.embedding IS NULL`);
+      conds.push(`cc.${staleCol} IS NULL`);
     }
     conds.push(`NOT (COALESCE(p.frontmatter, '{}'::jsonb) ? 'embed_skip')`);
     if (opts?.sourceId !== undefined) {
@@ -2394,7 +2408,7 @@ export class PGLiteEngine implements BrainEngine {
     return { where: conds.join(' AND '), params };
   }
 
-  async countStaleChunks(opts?: { sourceId?: string; signature?: string }): Promise<number> {
+  async countStaleChunks(opts?: { sourceId?: string; signature?: string; embeddingColumn?: ResolvedColumn }): Promise<number> {
     // D7: source-scoped count for `gbrain embed --stale --source X`. Always
     // JOIN pages so embed-skip + signature predicates apply. PGLite is
     // PostgreSQL 17.5 in WASM and supports the full JSONB operator set.
@@ -2410,7 +2424,7 @@ export class PGLiteEngine implements BrainEngine {
     return Number(count);
   }
 
-  async sumStaleChunkChars(opts?: { sourceId?: string; signature?: string }): Promise<number> {
+  async sumStaleChunkChars(opts?: { sourceId?: string; signature?: string; embeddingColumn?: ResolvedColumn }): Promise<number> {
     // Sibling of countStaleChunks: same stale predicate, summing chunk_text
     // length for the sync cost preview. ::bigint guards int4 overflow.
     const { where, params } = this.buildStaleChunkWhere(opts);
@@ -2463,11 +2477,17 @@ export class PGLiteEngine implements BrainEngine {
     sourceId?: string;
     orderBy?: 'page_id' | 'updated_desc';
     afterUpdatedAt?: string | null;
+    embeddingColumn?: ResolvedColumn;
   }): Promise<StaleChunkRow[]> {
     const limit = opts?.batchSize ?? 2000;
     const afterPid = opts?.afterPageId ?? 0;
     const afterIdx = opts?.afterChunkIndex ?? -1;
     const orderBy = opts?.orderBy ?? 'page_id';
+    // #1262: staleness follows the caller-resolved write column (validated +
+    // quoted identifier); legacy `embedding` otherwise.
+    const staleCol = opts?.embeddingColumn
+      ? buildVectorCastFragment(opts.embeddingColumn).col
+      : 'embedding';
 
     // v0.41.18.0 (A13, codex #9): --priority recent path. See postgres-engine
     // sibling for full rationale. Same composite cursor + ORDER BY.
@@ -2481,7 +2501,7 @@ export class PGLiteEngine implements BrainEngine {
                   p.updated_at
              FROM content_chunks cc
              JOIN pages p ON p.id = cc.page_id
-            WHERE cc.embedding IS NULL
+            WHERE cc.${staleCol} IS NULL
               AND NOT (COALESCE(p.frontmatter, '{}'::jsonb) ? 'embed_skip')
             ORDER BY p.updated_at DESC NULLS LAST, p.id ASC, cc.chunk_index ASC
             LIMIT $1`,
@@ -2492,7 +2512,7 @@ export class PGLiteEngine implements BrainEngine {
                   p.updated_at
              FROM content_chunks cc
              JOIN pages p ON p.id = cc.page_id
-            WHERE cc.embedding IS NULL
+            WHERE cc.${staleCol} IS NULL
               AND NOT (COALESCE(p.frontmatter, '{}'::jsonb) ? 'embed_skip')
               AND (
                 p.updated_at < $1::timestamptz
@@ -2511,7 +2531,7 @@ export class PGLiteEngine implements BrainEngine {
                 p.updated_at
            FROM content_chunks cc
            JOIN pages p ON p.id = cc.page_id
-          WHERE cc.embedding IS NULL
+          WHERE cc.${staleCol} IS NULL
             AND p.source_id = $1
             AND NOT (COALESCE(p.frontmatter, '{}'::jsonb) ? 'embed_skip')
           ORDER BY p.updated_at DESC NULLS LAST, p.id ASC, cc.chunk_index ASC
@@ -2523,7 +2543,7 @@ export class PGLiteEngine implements BrainEngine {
                 p.updated_at
            FROM content_chunks cc
            JOIN pages p ON p.id = cc.page_id
-          WHERE cc.embedding IS NULL
+          WHERE cc.${staleCol} IS NULL
             AND p.source_id = $1
             AND NOT (COALESCE(p.frontmatter, '{}'::jsonb) ? 'embed_skip')
             AND (
@@ -2548,7 +2568,7 @@ export class PGLiteEngine implements BrainEngine {
                 cc.model, cc.token_count, p.source_id, cc.page_id
            FROM content_chunks cc
            JOIN pages p ON p.id = cc.page_id
-          WHERE cc.embedding IS NULL
+          WHERE cc.${staleCol} IS NULL
             AND NOT (COALESCE(p.frontmatter, '{}'::jsonb) ? 'embed_skip')
             AND (cc.page_id, cc.chunk_index) > ($1, $2)
           ORDER BY cc.page_id, cc.chunk_index
@@ -2562,7 +2582,7 @@ export class PGLiteEngine implements BrainEngine {
               cc.model, cc.token_count, p.source_id, cc.page_id
          FROM content_chunks cc
          JOIN pages p ON p.id = cc.page_id
-        WHERE cc.embedding IS NULL
+        WHERE cc.${staleCol} IS NULL
           AND p.source_id = $1
           AND NOT (COALESCE(p.frontmatter, '{}'::jsonb) ? 'embed_skip')
           AND (cc.page_id, cc.chunk_index) > ($2, $3)

src/core/postgres-engine.ts

@@ -50,6 +50,7 @@ import type {
   BrainStats, BrainHealth,
   IngestLogEntry, IngestLogInput,
   EngineConfig,
+  ResolvedColumn,
   EvalCandidate, EvalCandidateInput,
   EvalCaptureFailure, EvalCaptureFailureReason,
   SalienceOpts, SalienceResult, AnomaliesOpts, AnomalyResult,
@@ -2380,13 +2381,21 @@ export class PostgresEngine implements BrainEngine {
   }
 
   // Chunks
-  async upsertChunks(slug: string, chunks: ChunkInput[], opts?: { sourceId?: string } & BatchOpts): Promise<void> {
+  async upsertChunks(slug: string, chunks: ChunkInput[], opts?: { sourceId?: string; embeddingColumn?: ResolvedColumn } & BatchOpts): Promise<void> {
     return this.batchRetry(opts?.auditSite ?? 'upsertChunks', opts?.signal, () => this._upsertChunksOnce(slug, chunks, opts), chunks.length);
   }
 
-  private async _upsertChunksOnce(slug: string, chunks: ChunkInput[], opts?: { sourceId?: string }): Promise<void> {
+  private async _upsertChunksOnce(slug: string, chunks: ChunkInput[], opts?: { sourceId?: string; embeddingColumn?: ResolvedColumn }): Promise<void> {
     const sql = this.sql;
     const sourceId = opts?.sourceId ?? 'default';
+    // #1262: caller-resolved write target for TEXT embeddings. Descriptor
+    // names are identifier-validated + quoted by buildVectorCastFragment;
+    // omitted => legacy `embedding vector`.
+    const targetFragment = opts?.embeddingColumn
+      ? buildVectorCastFragment(opts.embeddingColumn)
+      : undefined;
+    const targetCol = targetFragment?.col ?? 'embedding';
+    const embeddingCast = targetFragment?.castSql.replace('$1::', '') ?? 'vector';
 
     // Source-scope the page-id lookup. Without this filter, multi-source
     // brains where the slug exists in 2+ sources return >1 row and the
@@ -2413,7 +2422,7 @@ export class PostgresEngine implements BrainEngine {
     // scope metadata through upserts.
     // v0.27.1 (Phase 8): added `modality` + `embedding_image` to the column
     // list. Image chunks pass embedding=null + embedding_image=Float32Array.
-    const cols = '(page_id, chunk_index, chunk_text, chunk_source, embedding, model, token_count, embedded_at, language, symbol_name, symbol_type, start_line, end_line, parent_symbol_path, doc_comment, symbol_name_qualified, modality, embedding_image)';
+    const cols = `(page_id, chunk_index, chunk_text, chunk_source, ${targetCol}, model, token_count, embedded_at, language, symbol_name, symbol_type, start_line, end_line, parent_symbol_path, doc_comment, symbol_name_qualified, modality, embedding_image)`;
     const rows: string[] = [];
     const params: unknown[] = [];
     let paramIdx = 1;
@@ -2430,7 +2439,7 @@ export class PostgresEngine implements BrainEngine {
         : null;
       const modality = chunk.modality ?? 'text';
 
-      const embeddingPh = embeddingStr ? `$${paramIdx++}::vector` : 'NULL';
+      const embeddingPh = embeddingStr ? `$${paramIdx++}::${embeddingCast}` : 'NULL';
       const embeddedAtPh = embeddingStr ? 'now()' : 'NULL';
       const embeddingImagePh = embeddingImageStr ? `$${paramIdx++}::vector` : 'NULL';
 
@@ -2478,19 +2487,19 @@ export class PostgresEngine implements BrainEngine {
        ON CONFLICT (page_id, chunk_index) DO UPDATE SET
          chunk_text = EXCLUDED.chunk_text,
          chunk_source = EXCLUDED.chunk_source,
-         embedding = CASE
-           WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.embedding
-           WHEN content_chunks.embedding IS NULL THEN EXCLUDED.embedding
+         ${targetCol} = CASE
+           WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.${targetCol}
+           WHEN content_chunks.${targetCol} IS NULL THEN EXCLUDED.${targetCol}
            WHEN EXCLUDED.embedded_at IS NOT NULL
                 AND (content_chunks.embedded_at IS NULL OR EXCLUDED.embedded_at > content_chunks.embedded_at)
-                THEN EXCLUDED.embedding
-           ELSE content_chunks.embedding
+                THEN EXCLUDED.${targetCol}
+           ELSE content_chunks.${targetCol}
          END,
          model = COALESCE(EXCLUDED.model, content_chunks.model),
          token_count = EXCLUDED.token_count,
          embedded_at = CASE
-           WHEN EXCLUDED.chunk_text != content_chunks.chunk_text AND EXCLUDED.embedding IS NULL THEN NULL
-           WHEN content_chunks.embedding IS NULL AND EXCLUDED.embedding IS NOT NULL THEN EXCLUDED.embedded_at
+           WHEN EXCLUDED.chunk_text != content_chunks.chunk_text AND EXCLUDED.${targetCol} IS NULL THEN NULL
+           WHEN content_chunks.${targetCol} IS NULL AND EXCLUDED.${targetCol} IS NOT NULL THEN EXCLUDED.embedded_at
            WHEN EXCLUDED.embedded_at IS NOT NULL
                 AND (content_chunks.embedded_at IS NULL OR EXCLUDED.embedded_at > content_chunks.embedded_at)
                 THEN EXCLUDED.embedded_at
@@ -2530,14 +2539,19 @@ export class PostgresEngine implements BrainEngine {
    * embedding_signature drift (NULL grandfathered). Shared by
    * countStaleChunks + sumStaleChunkChars (parity with the PGLite sibling).
    */
-  private buildStaleChunkWhere(opts?: { sourceId?: string; signature?: string }): { where: string; params: unknown[] } {
+  private buildStaleChunkWhere(opts?: { sourceId?: string; signature?: string; embeddingColumn?: ResolvedColumn }): { where: string; params: unknown[] } {
+    // #1262: staleness targets the caller-resolved write column when set
+    // (identifier-validated + quoted); legacy `embedding` otherwise.
+    const staleCol = opts?.embeddingColumn
+      ? buildVectorCastFragment(opts.embeddingColumn).col
+      : 'embedding';
     const params: unknown[] = [];
     const conds: string[] = [];
     if (opts?.signature !== undefined) {
       params.push(opts.signature);
-      conds.push(`(cc.embedding IS NULL OR (p.embedding_signature IS NOT NULL AND p.embedding_signature <> $${params.length}))`);
+      conds.push(`(cc.${staleCol} IS NULL OR (p.embedding_signature IS NOT NULL AND p.embedding_signature <> $${params.length}))`);
     } else {
-      conds.push(`cc.embedding IS NULL`);
+      conds.push(`cc.${staleCol} IS NULL`);
     }
     conds.push(`NOT (COALESCE(p.frontmatter, '{}'::jsonb) ? 'embed_skip')`);
     if (opts?.sourceId !== undefined) {
@@ -2547,7 +2561,7 @@ export class PostgresEngine implements BrainEngine {
     return { where: conds.join(' AND '), params };
   }
 
-  async countStaleChunks(opts?: { sourceId?: string; signature?: string }): Promise<number> {
+  async countStaleChunks(opts?: { sourceId?: string; signature?: string; embeddingColumn?: ResolvedColumn }): Promise<number> {
     // Always JOIN pages so the embed_skip + signature predicates apply.
     // D7: source_id scoping. v0.41.31: optional signature widens staleness
     // to embedding_signature drift (NULL grandfathered).
@@ -2565,7 +2579,7 @@ export class PostgresEngine implements BrainEngine {
     });
   }
 
-  async sumStaleChunkChars(opts?: { sourceId?: string; signature?: string }): Promise<number> {
+  async sumStaleChunkChars(opts?: { sourceId?: string; signature?: string; embeddingColumn?: ResolvedColumn }): Promise<number> {
     // Sibling of countStaleChunks: same stale predicate, summing chunk_text
     // length for the sync cost preview. ::bigint guards int4 overflow.
     const { where, params } = this.buildStaleChunkWhere(opts);
@@ -2618,11 +2632,18 @@ export class PostgresEngine implements BrainEngine {
     sourceId?: string;
     orderBy?: 'page_id' | 'updated_desc';
     afterUpdatedAt?: string | null;
+    embeddingColumn?: ResolvedColumn;
   }): Promise<StaleChunkRow[]> {
     const limit = opts?.batchSize ?? 2000;
     const afterPid = opts?.afterPageId ?? 0;
     const afterIdx = opts?.afterChunkIndex ?? -1;
     const orderBy = opts?.orderBy ?? 'page_id';
+    // #1262: staleness follows the caller-resolved write column (validated +
+    // quoted identifier); legacy `embedding` otherwise. Interpolated below as
+    // an unsafe FRAGMENT (identifiers can't be bound parameters).
+    const staleCol = opts?.embeddingColumn
+      ? buildVectorCastFragment(opts.embeddingColumn).col
+      : 'embedding';
 
     // RLS scope binding (opt-in via GBRAIN_RLS_SCOPE_BINDING).
     return await this.withScopedReadTransaction(undefined, opts?.sourceId, async (tx) => {
@@ -2639,7 +2660,7 @@ export class PostgresEngine implements BrainEngine {
                    p.updated_at
             FROM content_chunks cc
             JOIN pages p ON p.id = cc.page_id
-            WHERE cc.embedding IS NULL
+            WHERE ${tx.unsafe(`cc.${staleCol} IS NULL`)}
               AND NOT (COALESCE(p.frontmatter, '{}'::jsonb) ? 'embed_skip')
             ORDER BY p.updated_at DESC NULLS LAST, p.id ASC, cc.chunk_index ASC
             LIMIT ${limit}
@@ -2649,7 +2670,7 @@ export class PostgresEngine implements BrainEngine {
                    p.updated_at
             FROM content_chunks cc
             JOIN pages p ON p.id = cc.page_id
-            WHERE cc.embedding IS NULL
+            WHERE ${tx.unsafe(`cc.${staleCol} IS NULL`)}
               AND NOT (COALESCE(p.frontmatter, '{}'::jsonb) ? 'embed_skip')
               AND (
                 p.updated_at < ${afterUpdated}::timestamptz
@@ -2667,7 +2688,7 @@ export class PostgresEngine implements BrainEngine {
                  p.updated_at
           FROM content_chunks cc
           JOIN pages p ON p.id = cc.page_id
-          WHERE cc.embedding IS NULL
+          WHERE ${tx.unsafe(`cc.${staleCol} IS NULL`)}
             AND p.source_id = ${opts.sourceId}
             AND NOT (COALESCE(p.frontmatter, '{}'::jsonb) ? 'embed_skip')
           ORDER BY p.updated_at DESC NULLS LAST, p.id ASC, cc.chunk_index ASC
@@ -2678,7 +2699,7 @@ export class PostgresEngine implements BrainEngine {
                  p.updated_at
           FROM content_chunks cc
           JOIN pages p ON p.id = cc.page_id
-          WHERE cc.embedding IS NULL
+          WHERE ${tx.unsafe(`cc.${staleCol} IS NULL`)}
             AND p.source_id = ${opts.sourceId}
             AND NOT (COALESCE(p.frontmatter, '{}'::jsonb) ? 'embed_skip')
             AND (
@@ -2698,7 +2719,7 @@ export class PostgresEngine implements BrainEngine {
                  cc.model, cc.token_count, p.source_id, cc.page_id
           FROM content_chunks cc
           JOIN pages p ON p.id = cc.page_id
-          WHERE cc.embedding IS NULL
+          WHERE ${tx.unsafe(`cc.${staleCol} IS NULL`)}
             AND NOT (COALESCE(p.frontmatter, '{}'::jsonb) ? 'embed_skip')
             AND (cc.page_id, cc.chunk_index) > (${afterPid}, ${afterIdx})
           ORDER BY cc.page_id, cc.chunk_index
@@ -2711,7 +2732,7 @@ export class PostgresEngine implements BrainEngine {
                cc.model, cc.token_count, p.source_id, cc.page_id
         FROM content_chunks cc
         JOIN pages p ON p.id = cc.page_id
-        WHERE cc.embedding IS NULL
+        WHERE ${tx.unsafe(`cc.${staleCol} IS NULL`)}
           AND p.source_id = ${opts.sourceId}
           AND NOT (COALESCE(p.frontmatter, '{}'::jsonb) ? 'embed_skip')
           AND (cc.page_id, cc.chunk_index) > (${afterPid}, ${afterIdx})

src/core/search/embedding-column.ts

@@ -443,6 +443,80 @@ export function resolveEmbeddingColumn(
   };
 }
 
+/**
+ * Resolves the WRITE-side embedding column for the currently configured
+ * embedding model (#1262). The read-side resolver above answers "which
+ * column does this query search?"; this one answers "which column should
+ * newly produced text embeddings land in?".
+ *
+ * Unlike read-side search, writes take no per-call column override. The
+ * import/embed boundary resolves once from merged config + gateway state
+ * and passes the descriptor into `engine.upsertChunks`; engines stay
+ * config-free (same contract as the read-side descriptor).
+ *
+ * Behavior:
+ *   - no user-declared `embedding_columns` => undefined (legacy brain,
+ *     writes keep targeting the default `embedding` column)
+ *   - a user-declared entry whose `provider` matches the current
+ *     embedding model => that entry's descriptor
+ *   - no provider match => undefined (fall back to legacy `embedding`)
+ *
+ * Only USER-declared entries are consulted — never the cfg-derived
+ * builtins. The `embedding_image` builtin's provider is the multimodal
+ * model; matching it here would misroute text embeddings into the image
+ * column. The no-match fallback is intentional: switching models before
+ * registering a matching column must not silently write vectors into an
+ * arbitrary column.
+ */
+export function resolveWriteColumn(cfg: GBrainConfig): ResolvedColumn | undefined {
+  const userColumns = cfg.embedding_columns;
+  if (
+    !userColumns ||
+    typeof userColumns !== 'object' ||
+    Array.isArray(userColumns) ||
+    Object.keys(userColumns).length === 0
+  ) {
+    return undefined;
+  }
+
+  // Same model-resolution chain as the registry builtin: cfg > gateway > default.
+  let gwModel: string | undefined;
+  try {
+    const gw = require('../ai/gateway.ts') as typeof import('../ai/gateway.ts');
+    gwModel = gw.getEmbeddingModel();
+  } catch {
+    // Gateway unconfigured — fall through to the canonical default.
+  }
+  const currentModel = cfg.embedding_model ?? gwModel ?? DEFAULT_EMBEDDING_MODEL;
+
+  for (const [name, entry] of Object.entries(userColumns)) {
+    if (!entry) continue;
+    validateColumnKey(name);
+    validateColumnConfig(name, entry);
+    if (entry.provider !== currentModel) continue;
+    return {
+      name,
+      type: entry.type,
+      dimensions: entry.dimensions,
+      embeddingModel: entry.provider,
+    };
+  }
+  return undefined;
+}
+
+/**
+ * Engine-boundary convenience: merged config (file/env + DB plane) →
+ * resolveWriteColumn. Dynamic import keeps config.ts out of this module's
+ * static graph (mirrors the gateway require above).
+ */
+export async function resolveWriteColumnForEngine(
+  engine: { getConfig(key: string): Promise<string | null | undefined> },
+): Promise<ResolvedColumn | undefined> {
+  const { loadConfigWithEngine } = await import('../config.ts');
+  const cfg = await loadConfigWithEngine(engine);
+  return cfg ? resolveWriteColumn(cfg) : undefined;
+}
+
 /**
  * True when the resolved column is the default `embedding` name.
  * Name-based check; does not compare embedding space.

test/chunker-version-gate.test.ts

@@ -15,22 +15,19 @@ import { describe, test, expect } from 'bun:test';
 import { CHUNKER_VERSION } from '../src/core/chunkers/code.ts';
 
 describe('Layer 12 — CHUNKER_VERSION constant', () => {
-  test('bumped to 5 for the estimated-token hard cap', () => {
+  test('bumped to 4 for Cathedral II', () => {
     // v3: v0.19.0 Chonkie parity (tokenizer + small-sibling merge).
     // v4: v0.20.0 Cathedral II (qualified names + parent scope + doc_comment
     //     + fence extraction + chunk-grain FTS). Folded into content_hash
     //     so any bump forces clean re-chunks on next sync.
-    // v5: estimated-token hard cap on AST-path chunks (capCodeChunks) so
-    //     un-subdividable giant nodes can't overflow strict embedding
-    //     server token limits.
-    expect(CHUNKER_VERSION).toBe(5);
+    expect(CHUNKER_VERSION).toBe(4);
   });
 
   test('is stable across imports (not recomputed at call time)', async () => {
     const a = (await import('../src/core/chunkers/code.ts')).CHUNKER_VERSION;
     const b = (await import('../src/core/chunkers/code.ts')).CHUNKER_VERSION;
     expect(a).toBe(b);
-    expect(a).toBe(5);
+    expect(a).toBe(4);
   });
 });
 

test/chunkers/code.test.ts

@@ -10,8 +10,8 @@ import { describe, test, expect } from 'bun:test';
 import { chunkCodeText, detectCodeLanguage, CHUNKER_VERSION } from '../../src/core/chunkers/code.ts';
 
 describe('CHUNKER_VERSION', () => {
-  test('v5: estimated-token hard cap on AST-path chunks', () => {
-    expect(CHUNKER_VERSION).toBe(5);
+  test('v0.20.0 Cathedral II Layer 12 bumped to 4', () => {
+    expect(CHUNKER_VERSION).toBe(4);
   });
 });
 

test/chunkers/recursive.test.ts

@@ -135,14 +135,13 @@ describe('Recursive Text Chunker', () => {
 });
 
 describe('CJK chunking (v0.32.7)', () => {
-  test('MARKDOWN_CHUNKER_VERSION is 4', async () => {
+  test('MARKDOWN_CHUNKER_VERSION is 3', async () => {
     // v0.40.3.0: bumped 2→3 to signal the post-upgrade reembed sweep that
-    // contextual retrieval wrapping is now applied at embed time.
-    // v4: estimated-token hard cap + whitespace-word undercount floor
-    // (URL-dense docs produced chunks past strict embedding server token
-    // limits). Boundary change → forces re-chunk for chunker_version < 4.
+    // contextual retrieval wrapping is now applied at embed time. Chunk
+    // boundaries themselves are unchanged; the bump forces re-embed for
+    // pages where chunker_version < 3.
     const mod = await import('../../src/core/chunkers/recursive.ts');
-    expect(mod.MARKDOWN_CHUNKER_VERSION).toBe(4);
+    expect(mod.MARKDOWN_CHUNKER_VERSION).toBe(3);
   });
 
   test('long pure-Chinese paragraph splits into multiple chunks', () => {

test/chunkers/token-cap.test.ts

@@ -1,195 +0,0 @@
-/**
- * Markdown chunker v4 / code chunker v5 — estimated-token hard cap
- * regression tests.
- *
- * Reproduces a field failure: a local llama-server embedding backend
- * (`-ub 2048`) crashes deterministically (trace/BPT trap → EOF at the
- * client) when a single chunk exceeds ~2,050 real tokens. Two content
- * shapes triggered it:
- *
- *   1. Korean docs carrying one long source URL per line.
- *      The URLs' ASCII mass pushes CJK density below 0.30, flipping
- *      countCJKAwareWords to whitespace counting, where a 150-char URL
- *      counts as ONE word → chunks ballooned to 3-4K chars ≈ 2,000+
- *      real tokens (URL soup tokenizes at ~1.6 chars/token).
- *
- *   2. Large JSON code blocks (~7K chars) that the word pipeline
- *      undercounts the same way (few whitespace tokens).
- *
- * The fix: every emitted chunk must satisfy
- *   estimateEmbeddingTokens(chunk) <= maxTokens (default 1500)
- * where the estimate deliberately OVERSTATES real tokenizer counts.
- */
-
-import { describe, test, expect } from 'bun:test';
-import { chunkText, capByEstimatedTokens, DEFAULT_MAX_EST_TOKENS } from '../../src/core/chunkers/recursive.ts';
-import { chunkCodeText } from '../../src/core/chunkers/code.ts';
-import { estimateEmbeddingTokens } from '../../src/core/cjk.ts';
-
-/** Synthesize the failing shape: Korean rollup lines each ending in a long Notion URL. */
-function urlDenseKoreanRollup(lines: number): string {
-  const out: string[] = ['# 링크가 줄마다 붙는 한국어 예시 문서', ''];
-  for (let i = 0; i < lines; i++) {
-    const hex32 = (i * 2654435761 >>> 0).toString(16).padStart(8, '0').repeat(4);
-    out.push(
-      `- **항목 ${i}**: 이 줄은 청커 동작 검증을 위한 의미 없는 한국어 예시 문장입니다 · 전화 000-0000-${String(1000 + i)} · ` +
-      `이메일 user${i}@example.com · 링크: https://docs.example.com/pages/${hex32}?v=abcdef0123456789&ref=sample`,
-    );
-  }
-  return out.join('\n');
-}
-
-/** Synthesize a large pretty-printed JSON block with CJK values. */
-function bigJsonBlock(targetChars: number): string {
-  const entries: string[] = [];
-  let i = 0;
-  let len = 0;
-  while (len < targetChars) {
-    const row =
-      `  "item_${i}": { "name": "예시-${i}", "url": "https://example.com/api/v2/items/${i}?token=abc${i}def", "qty": ${i % 100}, "memo": "한국어 값이 섞인 예시 데이터" }`;
-    entries.push(row);
-    len += row.length;
-    i++;
-  }
-  return `{\n${entries.join(',\n')}\n}`;
-}
-
-describe('v4 estimated-token cap — URL-dense Korean doc (field-failure shape)', () => {
-  test('every chunk stays under the estimated-token cap', () => {
-    const md = urlDenseKoreanRollup(60);
-    const chunks = chunkText(md);
-    expect(chunks.length).toBeGreaterThan(0);
-    for (const c of chunks) {
-      expect(estimateEmbeddingTokens(c.text)).toBeLessThanOrEqual(DEFAULT_MAX_EST_TOKENS);
-    }
-  });
-
-  test('no chunk reaches the measured 3K-char danger zone for URL soup', () => {
-    const md = urlDenseKoreanRollup(60);
-    const chunks = chunkText(md);
-    // 1500 est tokens at the OTHER weight (0.75/char) bounds chunks to
-    // ~2,000 chars for pure ASCII — well under the ~3,300 chars where
-    // URL-dense content crosses ~2,050 real tokens (1.6 chars/token).
-    for (const c of chunks) {
-      expect(c.text.length).toBeLessThanOrEqual(2600);
-    }
-  });
-
-  test('content is preserved (no lines dropped by the cap)', () => {
-    const md = urlDenseKoreanRollup(60);
-    const chunks = chunkText(md);
-    const joined = chunks.map((c) => c.text).join('\n');
-    // Spot-check first / middle / last rollup lines survive chunking.
-    for (const marker of ['항목 0', '항목 30', '항목 59']) {
-      expect(joined).toContain(marker);
-    }
-  });
-});
-
-describe('v4 estimated-token cap — large JSON blocks', () => {
-  test('7K-char pretty JSON through the prose path stays under the cap', () => {
-    const md = `설정 파일 원문 보존:\n\n\`\`\`\n${bigJsonBlock(7000)}\n\`\`\`\n`;
-    const chunks = chunkText(md);
-    expect(chunks.length).toBeGreaterThan(1);
-    for (const c of chunks) {
-      expect(estimateEmbeddingTokens(c.text)).toBeLessThanOrEqual(DEFAULT_MAX_EST_TOKENS);
-    }
-  });
-
-  test('7K-char minified JSON (single whitespace-less token) stays under the cap', () => {
-    const minified = bigJsonBlock(7000).replace(/\n\s*/g, '');
-    const chunks = chunkText(minified);
-    expect(chunks.length).toBeGreaterThan(1);
-    for (const c of chunks) {
-      expect(estimateEmbeddingTokens(c.text)).toBeLessThanOrEqual(DEFAULT_MAX_EST_TOKENS);
-    }
-  });
-
-  test('json fence via the code chunker stays under the cap (+header slack)', async () => {
-    const chunks = await chunkCodeText(bigJsonBlock(7000), 'fence.json');
-    expect(chunks.length).toBeGreaterThan(0);
-    for (const c of chunks) {
-      // buildChunk prepends a short "[JSON] fence.json:…" header AFTER the
-      // body-level cap; allow ~60 est tokens of header slack. Real-token
-      // safety margin (2,050 − overestimated 1,500) absorbs this easily.
-      expect(estimateEmbeddingTokens(c.text)).toBeLessThanOrEqual(DEFAULT_MAX_EST_TOKENS + 60);
-    }
-  });
-});
-
-describe('v4 word-count floor — behavior preserved for normal content', () => {
-  test('Latin prose chunking is unchanged by the floor (avg word < 6 chars)', () => {
-    const prose = Array.from({ length: 120 }, (_, i) =>
-      `This is sentence number ${i} and it talks about ordinary things in plain words.`,
-    ).join(' ');
-    const chunks = chunkText(prose);
-    // Historical behavior: ~1,560 whitespace words → multiple ~300-word chunks.
-    expect(chunks.length).toBeGreaterThan(3);
-    for (const c of chunks) {
-      const words = c.text.split(/\s+/).length;
-      expect(words).toBeLessThanOrEqual(300 * 1.5 + 50); // merge cap + overlap
-    }
-  });
-
-  test('Korean prose (CJK-dense, no URLs) never triggers the token cap', () => {
-    const prose = Array.from({ length: 80 }, (_, i) =>
-      `이 문장은 순수 한국어 산문의 청킹 동작을 확인하기 위한 ${i}번째 예시 문장입니다.`,
-    ).join(' ');
-    const chunks = chunkText(prose);
-    expect(chunks.length).toBeGreaterThan(1);
-    for (const c of chunks) {
-      // CJK-dense chunks are char-counted (≈450 max) — nowhere near 1500.
-      expect(estimateEmbeddingTokens(c.text)).toBeLessThanOrEqual(700);
-    }
-  });
-});
-
-describe('capByEstimatedTokens unit behavior', () => {
-  test('returns input unchanged when under the cap', () => {
-    expect(capByEstimatedTokens('short text', 1500)).toEqual(['short text']);
-    expect(capByEstimatedTokens('', 1500)).toEqual([]);
-  });
-
-  test('prefers newline cut points within the lookback window', () => {
-    const line = 'x'.repeat(100);
-    const text = Array.from({ length: 40 }, () => line).join('\n');
-    const pieces = capByEstimatedTokens(text, 1000);
-    expect(pieces.length).toBeGreaterThan(1);
-    for (const p of pieces) {
-      // Every piece should be whole lines (multiples of the 100-char line).
-      for (const l of p.split('\n')) {
-        expect(l).toBe(line);
-      }
-    }
-  });
-
-  test('makes forward progress on whitespace-less input (hard cut)', () => {
-    const blob = 'a'.repeat(10_000);
-    const pieces = capByEstimatedTokens(blob, 1000);
-    expect(pieces.length).toBeGreaterThan(1);
-    expect(pieces.join('')).toBe(blob);
-    for (const p of pieces) {
-      expect(estimateEmbeddingTokens(p)).toBeLessThanOrEqual(1000);
-    }
-  });
-});
-
-describe('estimateEmbeddingTokens — weight sanity', () => {
-  test('overestimates URL-dense ASCII (0.75/char ≥ measured ~0.63/char)', () => {
-    const url = 'https://docs.example.com/pages/a1b2c3d4e5f6a1b2c3d4e5f6a1b2c3d4?v=abc&ref=sample';
-    const est = estimateEmbeddingTokens(url);
-    expect(est).toBeGreaterThanOrEqual(Math.floor(url.length * 0.7));
-  });
-
-  test('counts CJK at 1 token/char', () => {
-    expect(estimateEmbeddingTokens('가나다라마')).toBe(5);
-  });
-
-  test('whitespace is nearly free', () => {
-    expect(estimateEmbeddingTokens('   \n\t  ')).toBeLessThanOrEqual(1);
-  });
-
-  test('empty string is 0', () => {
-    expect(estimateEmbeddingTokens('')).toBe(0);
-  });
-});

test/e2e/embedding-column-pglite.test.ts

@@ -241,3 +241,136 @@ describe('buildVectorCastFragment — engine SQL composer (D3)', () => {
     expect(castSql).toBe('$1::halfvec(2560)');
   });
 });
+
+describe('PGLite engine: upsertChunks write-side ResolvedColumn descriptor (#1262)', () => {
+  test('halfvec descriptor writes the text embedding to the alternate column, not legacy embedding', async () => {
+    await engine.putPage('docs/write-alt-pglite', {
+      type: 'concept',
+      title: 'Write alt column PGLite',
+      compiled_truth: 'PGLite write-side alternate embedding column test.',
+    });
+
+    const descriptor: ResolvedColumn = {
+      name: 'embedding_ze',
+      type: 'halfvec',
+      dimensions: 2560,
+      embeddingModel: 'zeroentropyai:zembed-1',
+    };
+    await engine.upsertChunks('docs/write-alt-pglite', [
+      {
+        chunk_index: 0,
+        chunk_text: 'PGLite write-side alternate embedding column test.',
+        chunk_source: 'compiled_truth',
+        embedding: new Float32Array(2560).fill(0.25),
+      },
+    ], { embeddingColumn: descriptor });
+
+    const rows = await engine.executeRaw<{
+      has_default: boolean;
+      has_ze: boolean;
+      has_embedded_at: boolean;
+    }>(
+      `SELECT embedding IS NOT NULL AS has_default,
+              embedding_ze IS NOT NULL AS has_ze,
+              embedded_at IS NOT NULL AS has_embedded_at
+         FROM content_chunks cc
+         JOIN pages p ON p.id = cc.page_id
+        WHERE p.slug = 'docs/write-alt-pglite'`,
+    );
+    expect(rows.length).toBe(1);
+    expect(rows[0].has_default).toBe(false);
+    expect(rows[0].has_ze).toBe(true);
+    expect(rows[0].has_embedded_at).toBe(true);
+  });
+
+  test('text-unchanged re-upsert without a vector preserves the alternate-column embedding', async () => {
+    const descriptor: ResolvedColumn = {
+      name: 'embedding_ze',
+      type: 'halfvec',
+      dimensions: 2560,
+      embeddingModel: 'zeroentropyai:zembed-1',
+    };
+    // Same chunk_text, no embedding: the ON CONFLICT CASE must keep the
+    // existing alternate-column vector (D24 semantics follow the column).
+    await engine.upsertChunks('docs/write-alt-pglite', [
+      {
+        chunk_index: 0,
+        chunk_text: 'PGLite write-side alternate embedding column test.',
+        chunk_source: 'compiled_truth',
+      },
+    ], { embeddingColumn: descriptor });
+    const rows = await engine.executeRaw<{ has_ze: boolean }>(
+      `SELECT embedding_ze IS NOT NULL AS has_ze
+         FROM content_chunks cc
+         JOIN pages p ON p.id = cc.page_id
+        WHERE p.slug = 'docs/write-alt-pglite'`,
+    );
+    expect(rows).toEqual([{ has_ze: true }]);
+  });
+});
+
+describe('PGLite: embed --stale converges on an alt-column brain (#1262)', () => {
+  test('boundary resolves the write column; stale scan does not re-select embedded rows', async () => {
+    const { runEmbedCore } = await import('../../src/commands/embed.ts');
+    const local = new PGLiteEngine();
+    const previousHome = process.env.GBRAIN_HOME;
+    process.env.GBRAIN_HOME = `/tmp/gbrain-write-col-stale-${Date.now()}`;
+    try {
+      await local.connect({});
+      await local.initSchema();
+      await (local as any).db.exec(
+        `ALTER TABLE content_chunks ADD COLUMN IF NOT EXISTS embedding_ze halfvec(2560)`,
+      );
+
+      const descriptor: ResolvedColumn = {
+        name: 'embedding_ze',
+        type: 'halfvec',
+        dimensions: 2560,
+        embeddingModel: 'zeroentropyai:zembed-1',
+      };
+      await local.setConfig('embedding_columns', JSON.stringify({
+        embedding_ze: { provider: 'zeroentropyai:zembed-1', dimensions: 2560, type: 'halfvec' },
+      }));
+      configureGateway({
+        embedding_model: 'zeroentropyai:zembed-1',
+        embedding_dimensions: 2560,
+        env: {},
+      });
+
+      await local.putPage('docs/stale-alt-pglite', {
+        type: 'concept',
+        title: 'Dynamic stale column',
+        compiled_truth: 'A chunk that is embedded only in the dynamic column.',
+      });
+      await local.upsertChunks('docs/stale-alt-pglite', [
+        {
+          chunk_index: 0,
+          chunk_text: 'A chunk that is embedded only in the dynamic column.',
+          chunk_source: 'compiled_truth',
+          embedding: new Float32Array(2560).fill(0.25),
+        },
+      ], { embeddingColumn: descriptor });
+
+      // Engine-level contrast: legacy predicate still sees the row as stale;
+      // the alt-column predicate does not.
+      expect(await local.countStaleChunks()).toBe(1);
+      expect(await local.countStaleChunks({ embeddingColumn: descriptor })).toBe(0);
+      // sumStaleChunkChars feeds the sync cost gate — same predicate contract.
+      expect(await local.sumStaleChunkChars()).toBeGreaterThan(0);
+      expect(await local.sumStaleChunkChars({ embeddingColumn: descriptor })).toBe(0);
+      expect(await local.listStaleChunks({ embeddingColumn: descriptor, batchSize: 100 })).toHaveLength(0);
+      expect(await local.listStaleChunks({ batchSize: 100 })).toHaveLength(1);
+
+      // Boundary-level: `embed --stale --dry-run` resolves the write column
+      // from merged config + gateway and reports NOTHING to embed. Without
+      // the fix this reports 1 (perpetual re-embed loop).
+      const result = await runEmbedCore(local, { stale: true, dryRun: true });
+      expect(result.would_embed).toBe(0);
+    } finally {
+      await local.disconnect();
+      if (previousHome === undefined) delete process.env.GBRAIN_HOME;
+      else process.env.GBRAIN_HOME = previousHome;
+      resetGateway();
+    }
+  });
+});

test/e2e/embedding-column-postgres.test.ts

@@ -224,4 +224,54 @@ if (!dbUrl) {
       await engine.executeRaw(`UPDATE content_chunks SET embedding_voyage = '${v}'::vector WHERE id = ${dogId}`);
     });
   });
+
+  describe('Postgres: upsertChunks write-side ResolvedColumn descriptor (#1262)', () => {
+    const descriptor: ResolvedColumn = {
+      name: 'embedding_ze',
+      type: 'halfvec',
+      dimensions: 2560,
+      embeddingModel: 'zeroentropyai:zembed-1',
+    };
+
+    test('halfvec descriptor writes the text embedding to the alternate column, not legacy embedding', async () => {
+      await engine.putPage('docs/write-alt-postgres', {
+        type: 'concept',
+        title: 'Write alt column Postgres',
+        compiled_truth: 'Postgres write-side alternate embedding column test.',
+      });
+      await engine.upsertChunks('docs/write-alt-postgres', [
+        {
+          chunk_index: 0,
+          chunk_text: 'Postgres write-side alternate embedding column test.',
+          chunk_source: 'compiled_truth',
+          embedding: new Float32Array(2560).fill(0.25),
+        },
+      ], { embeddingColumn: descriptor });
+
+      const rows = await engine.executeRaw<{
+        has_default: boolean;
+        has_ze: boolean;
+      }>(
+        `SELECT embedding IS NOT NULL AS has_default,
+                embedding_ze IS NOT NULL AS has_ze
+           FROM content_chunks cc
+           JOIN pages p ON p.id = cc.page_id
+          WHERE p.slug = 'docs/write-alt-postgres'`,
+      );
+      expect(rows.length).toBe(1);
+      expect(rows[0].has_default).toBe(false);
+      expect(rows[0].has_ze).toBe(true);
+    }, 30_000);
+
+    test('stale scan follows the write-side column (count + list parity with the write target)', async () => {
+      // Legacy predicate: cat/dog/write-alt rows all have embedding NULL.
+      expect(await engine.countStaleChunks()).toBeGreaterThan(0);
+      // Alt-column predicate: every chunk has embedding_ze populated.
+      expect(await engine.countStaleChunks({ embeddingColumn: descriptor })).toBe(0);
+      expect(await engine.listStaleChunks({ embeddingColumn: descriptor, batchSize: 100 })).toHaveLength(0);
+      expect((await engine.listStaleChunks({ batchSize: 100 })).length).toBeGreaterThan(0);
+      // updated_desc arm uses the same predicate.
+      expect(await engine.listStaleChunks({ embeddingColumn: descriptor, orderBy: 'updated_desc', batchSize: 100 })).toHaveLength(0);
+    }, 30_000);
+  });
 }

test/search/embedding-column.serial.test.ts

@@ -13,9 +13,10 @@
  *     throw on unknown string.
  */
 
-import { describe, test, expect } from 'bun:test';
+import { describe, test, expect, afterAll, afterEach } from 'bun:test';
 import {
   resolveEmbeddingColumn,
+  resolveWriteColumn,
   getEmbeddingColumnRegistry,
   buildVectorCastFragment,
   quoteIdentifier,
@@ -34,6 +35,28 @@ import {
 } from '../../src/core/search/embedding-column.ts';
 import type { GBrainConfig } from '../../src/core/config.ts';
 import type { ResolvedColumn } from '../../src/core/types.ts';
+import { configureGateway, resetGateway } from '../../src/core/ai/gateway.ts';
+
+/**
+ * Teardown: reset AND re-apply the legacy preload config
+ * (test/helpers/legacy-embedding-preload.ts). A bare resetGateway() would
+ * leave the slot empty for the NEXT file's beforeAll (the preload's
+ * per-test beforeEach only fires before tests, not before beforeAll), which
+ * would make sibling PGLite fixtures initSchema at the 1280 default instead
+ * of the legacy 1536 their seed vectors assume.
+ */
+function restorePreloadGateway() {
+  resetGateway();
+  configureGateway({
+    embedding_model: 'openai:text-embedding-3-large',
+    embedding_dimensions: 1536,
+    env: { ...process.env },
+  });
+}
+
+afterAll(() => {
+  restorePreloadGateway();
+});
 
 function cfg(overrides: Partial<GBrainConfig> = {}): GBrainConfig {
   return { engine: 'pglite', ...overrides };
@@ -522,3 +545,89 @@ describe('codex /ship #4 — isCacheSafe (embedding-space-based skip)', () => {
     expect(isCacheSafe(r, cfg())).toBe(true);
   });
 });
+
+describe('resolveWriteColumn — write-side boundary resolution (#1262)', () => {
+  afterEach(() => {
+    restorePreloadGateway();
+  });
+
+  test('no registry / empty registry returns undefined (legacy single-column brain)', () => {
+    expect(resolveWriteColumn(cfg())).toBeUndefined();
+    expect(resolveWriteColumn(cfg({ embedding_columns: {} }))).toBeUndefined();
+  });
+
+  test('provider match via cfg.embedding_model returns the descriptor', () => {
+    const r = resolveWriteColumn(cfg({
+      embedding_model: 'voyage:voyage-3-large',
+      embedding_dimensions: 1024,
+      embedding_columns: {
+        embedding_voyage: { provider: 'voyage:voyage-3-large', dimensions: 1024, type: 'vector' },
+      },
+    }));
+    expect(r).toEqual({
+      name: 'embedding_voyage',
+      type: 'vector',
+      dimensions: 1024,
+      embeddingModel: 'voyage:voyage-3-large',
+    });
+  });
+
+  test('provider match via gateway state (cfg.embedding_model unset) returns descriptor', () => {
+    configureGateway({
+      embedding_model: 'zeroentropyai:zembed-1',
+      embedding_dimensions: 2560,
+      env: {},
+    });
+    const r = resolveWriteColumn(cfg({
+      embedding_columns: {
+        embedding_ze: { provider: 'zeroentropyai:zembed-1', dimensions: 2560, type: 'halfvec' },
+      },
+    }));
+    expect(r).toEqual({
+      name: 'embedding_ze',
+      type: 'halfvec',
+      dimensions: 2560,
+      embeddingModel: 'zeroentropyai:zembed-1',
+    });
+  });
+
+  test('no provider match returns undefined instead of guessing a column', () => {
+    configureGateway({
+      embedding_model: 'zeroentropyai:zembed-1',
+      embedding_dimensions: 2560,
+      env: {},
+    });
+    const r = resolveWriteColumn(cfg({
+      embedding_columns: {
+        embedding_voyage: { provider: 'voyage:voyage-3-large', dimensions: 1024, type: 'vector' },
+      },
+    }));
+    expect(r).toBeUndefined();
+  });
+
+  test('only USER-declared columns are consulted — multimodal builtin never captures text writes', () => {
+    // Current model equals the embedding_image BUILTIN's provider; a registry
+    // walk that consulted builtins would misroute text writes into the image
+    // column. resolveWriteColumn must return undefined here.
+    configureGateway({
+      embedding_model: 'voyage:voyage-multimodal-3',
+      embedding_dimensions: 1024,
+      env: {},
+    });
+    const r = resolveWriteColumn(cfg({
+      embedding_columns: {
+        embedding_other: { provider: 'openai:text-embedding-3-large', dimensions: 1536, type: 'vector' },
+      },
+    }));
+    expect(r).toBeUndefined();
+  });
+
+  test('malformed registry entry throws loud (same validation as the read side)', () => {
+    expect(() => resolveWriteColumn(cfg({
+      embedding_model: 'voyage:voyage-3-large',
+      embedding_columns: {
+        'bad"col': { provider: 'voyage:voyage-3-large', dimensions: 1024, type: 'vector' },
+      } as never,
+    }))).toThrow(EmbeddingColumnConfigError);
+  });
+});