Retune local embedding context defaults for Qwen3 ingestion

Context

The operator flagged that localModels.embedding.contextLimitTokens = 8192 and safeProcessingLimitTokens = 6144 are too small for real file-scale embedding workloads, especially when a raw source file can be thousands of lines. PR #12274 was merged on 2026-06-01 but explicitly did not retune these config defaults.

Duplicate sweep:

• #12264 is open and now has a comment preserving that embedding defaults are unresolved config-SSOT scope.
• The open issue search for localModels embedding contextLimitTokens 8192 6144 only returned #12264.
• No current open PR in the live queue fixes the embedding defaults.

This ticket extracts the embedding-default work into its own implementable child so the intent is not buried under the chat/LMS preload title.

The Problem

The configured OpenAI-compatible embedding model default is text-embedding-qwen3-embedding-8b, while localModels.embedding.* still carries conservative placeholder values:

• contextLimitTokens: 8192
• safeProcessingLimitTokens: 6144

The config comment itself says Qwen3-8B embedding typically supports 32K context and that 8K is only a conservative floor pending validation. The same block still says no active consumer reads these values, but #12274 introduced role-keyed LMS context loading that can consume the embedding context length. Separately, KB/Memory Core embedding pipelines are real production consumers of embedding capacity, and raw external repo ingestion can send file-scale chunks through the embedding path.

A 6144-token safe band is therefore not an acceptable settled default for the current local embedding role. It can preemptively reject or underconfigure valid file-scale ingestion, and it encourages operators to debug model/provider failures that are actually config-default drift.

The Architectural Reality

• ai/config.template.mjs owns the role-keyed local-model context defaults under localModels.chat and localModels.embedding.
• openAiCompatible.embeddingModel defaults to text-embedding-qwen3-embedding-8b with vectorDimension: 4096.
• ProviderReadinessHelper.buildLmsContextLengthsMap() maps chat and embedding model IDs to role-specific context lengths for local OpenAI-compatible loading.
• TextEmbeddingService and KB ingestion are embedding-role consumers; they must not inherit chat-sized assumptions or obsolete placeholder comments.

The Fix

Validate the actual configured local embedding model capability and retune the embedding-role defaults to a safe, documented operational band. The likely target is a 32K-class embedding context with an explicit safe-processing band, but the implementation must let the validation result determine the final values.

Contract Ledger Matrix

Decision Record Impact

None. This is a config-SSOT correction aligned with #12264 and #12090; it does not amend an ADR.

Acceptance Criteria

• V-B-A confirms the context capability of the configured default embedding model (text-embedding-qwen3-embedding-8b) from the serving stack or primary model docs before changing the defaults.
• localModels.embedding.contextLimitTokens is retuned from 8192 to the validated embedding-role capacity.
• localModels.embedding.safeProcessingLimitTokens is retuned from 6144 to an explicit documented headroom band appropriate for that capacity.
• The obsolete comment claim that no active consumer reads these values is removed or replaced with accurate consumer mapping.
• Tests cover default values, env overrides, and at least one role-aware context map/guardrail consumer so the embedding role cannot silently regress to a tiny band.
• PR evidence includes a file-scale sample rationale: a multi-thousand-line source file must not be rejected solely by the old 6144-token placeholder band when the configured embedding model supports a larger window.

Out of Scope

• Changing chat model/context defaults.
• Changing vector dimensions.
• Changing provider selection or forcing all deployments onto OpenAI-compatible.
• Implementing native Ollama startup parity; tracked separately.

Avoided Traps

• Do not leave a magic small default just because operators can override it. Defaults are the path most agents and dev environments will exercise.
• Do not retune chat and embedding together unless evidence shows the same model-role requirement; these are separate roles.
• Do not cite Qwen3's likely 32K capacity without validating the active served model and server behavior.

Related

• Parent: #12264
• Delivered mandate: #12090
• Triggering merged PR that left this out of scope: #12274
• Sibling follow-up: #12285

Origin Session ID: d9b4a887-57aa-43aa-8cec-a260ec35ce12

Handoff Retrieval Hints: localModels embedding contextLimitTokens 8192 safeProcessingLimitTokens 6144 Qwen3 embedding defaults; #12264 embedding role config-SSOT comment; buildLmsContextLengthsMap embeddingContextLength