Support explicitly selected model-family subsets in setup-pstack

[simonbloom]

The current setup requires Fable, Sol, GLM, and Opus to all probe successfully before it writes any configuration. That makes a supported Sol + GLM installation impossible even though the README says core workflows can run with fewer models.\n\nAdd an explicit active-family selection to setup-pstack. Probe, persist, and smoke only the selected families; preserve fail-closed behavior for every selected family; require all documented role rows; and reject descriptors from inactive or unprobed families. Keep the four-family profile as the first-run default.\n\nImmediate user case: Codex parent with GPT-5.6-Sol max as the primary/final lane and Z.ai Coding Plan GLM-5.3-Flash max for implementation/exploration. Claude remains installed but is not a dependency, and Grok remains outside the configured matrix.

[simonbloom]

Implemented the active-family subset contract locally on feature/opencode-glm-adapter, using the explicitly approved existing GLM adapter working tree as the baseline. The source candidate keeps the complete Fable/Sol/GLM/Opus example as the first-run proposal, makes the active set explicit, derives the rerun proposal from the current sheet, asks efforts and probes only active families, requires explicit replacements when deactivating lanes, rejects inactive/unprobed descriptors, and smokes every active family with an independent cross-provider judge when available.

Source validation completed on the candidate: 163 passing, 0 failing; strict TypeScript checks passed; PSTACK_STATIC_ONLY=1 bash tests/skill-collision-repro.sh passed; git diff --check passed. Candidate binary diff checksum is c951d274af5973cf4c768001d46a5c6e06bd5bfc3f676397f287e9e87ceebda1. Installed-cache and live Sol+GLM verification are being completed before this is considered done. No commit, push, PR, merge, tag, or release was created.

[simonbloom]

Installed and live-verified the active-family subset candidate from a real Codex parent. Both existing 1.2.1 caches retain 52 skills, and only setup-pstack/SKILL.md, provider-dispatch.md, and model-matrix.test.ts were refreshed from the validated source candidate.

Configured the explicit active set codex:gpt-5.6-sol@max + opencode:zai-coding-plan/glm-5.3-flash@max. The native Sol max probe passed. The external OpenCode GLM max probe passed with the exact model and variant and pinned-argv evidence. The installed behavioral smoke passed its native Sol candidate, external GLM candidate, and a separate GLM cross-judge launched only after both candidates were terminal. The Codex sheet contains all 16 documented role rows, Why/Reflect inherit the parent, its bounded AGENTS block is byte-identical, and multi_agent is true. No inactive family was probed or persisted; no Claude sheet was created.

Model-sheet SHA-256: 34bd2c31deff0ef1d79ef28d884a3207ac3ac0e33d17f63556ab4a6f93210a5d. Evidence directory: /Users/simonbloom/.codex/pstack-setup-receipts/2026-08-31-sol-glm. No commit, push, PR, merge, tag, or release was created.