fix(agents): preserve 3-segment model IDs (lmstudio/qwen/model-name)

[Copilot]

Any model ID with 3+ slash-separated segments was unconditionally treated as provider/model/variant — the last segment was stripped and injected as the variant field. LM Studio model IDs like lmstudio/qwen/qwen3.6-35b-a3b were silently rewritten to model: "lmstudio/qwen" + variant: "qwen3.6-35b-a3b", causing ProviderModelNotFoundError at spawn time.

Changes

• src/agents/index.ts — adds KNOWN_VARIANT_VALUES = new Set(['low', 'medium', 'high', 'max', 'xhigh', 'thinking']). The auto-split in applyOverrides() now only fires when the trailing segment is a known reasoning-effort token:

  // Before — strips ANY third segment
  if (modelSegments.length >= 3) { … }
  
  // After — only strips known variant tokens
  const hasEmbeddedVariant =
    modelSegments.length >= 3 && KNOWN_VARIANT_VALUES.has(lastSegment);
  if (hasEmbeddedVariant) { … }

• src/config/schema.ts — updates the variant field comment to document the whitelist behaviour and the LM Studio counter-example.

• docs/configuration.md — updates the variant field description to list all six known tokens.

• docs/releases/v7.5.3.md — release notes for the fix.

• tests/unit/agents/factory.test.ts — adds three regression tests:

  • lmstudio/qwen/qwen3.6-35b-a3b is passed through unchanged
  • multiple LM Studio agents in one config all preserve their full paths
  • low, medium, high, max, xhigh, thinking still trigger the deprecation split (backward compat preserved), including a 4-segment model (gateway/ns/gpt-5.3-codex/medium)

Scope note

Commit b739e80c was cherry-picked from the already-merged PR #773, which broadened the variant allowlist with max and xhigh and added the 4-segment split test case.