Promote the GPT-6 Codex families to stock and put them in the first-run defaults

[thisguymartin]

Problem

#9 added codex:gpt-6-astra, codex:gpt-6-sol, and codex:gpt-6-luna as an "additional" matrix: selectable in /setup-pstack, absent from the first-run sheet, and fenced off by tests so the upstream-owned stock matrix stayed byte-identical. That split was a sync-hygiene choice, not a technical one. In practice it means a fresh install still defaults every Codex role to GPT-5.6 Sol and the docs call the GPT-6 families "optional", which is not how this fork wants to present them.

Decision

Make the three GPT-6 families stock. One model matrix, no "additional" tier, and a first-run sheet that uses them.

Proposed first-run role map

Role	Today	Proposed
feature, refactoring	grok:grok-4.6@xhigh	codex:gpt-6-sol@high
bug-fix	codex:gpt-5.6-sol@max	codex:gpt-6-sol@high
perf-issue	codex:gpt-5.6-sol@max	codex:gpt-6-sol@high
hillclimb	codex:gpt-5.6-sol@max	codex:gpt-6-sol@high
judgment and prose	claude:fable@max	unchanged
hardest tasks	claude:fable@max	unchanged
how explorer	grok:grok-4.6@xhigh	codex:gpt-6-luna@high
how explainer	claude:fable@max	unchanged
why / reflect roles	inherit-parent	unchanged
swarm workers	grok:grok-4.6@xhigh	codex:gpt-6-luna@high
arena runners	fable, gpt-5.6-sol, grok, opus	claude:fable@max, codex:gpt-6-astra@high, grok:grok-4.6@xhigh, claude:opus@xhigh
arena cross-judge pool	same four	same as arena runners
architect runners	same four	same as arena runners
interrogate reviewers	same four	same as arena runners

Reasoning: GPT-6 Sol is Codex's coding workhorse, so it takes the solo code-writing roles. Luna is the fast low-cost tier, so it takes exploration and swarm verification lanes. Astra is the frontier tier, so it replaces GPT-5.6 Sol in every panel. codex:gpt-5.6-sol stays in the matrix as a selectable family for existing sheets; it just leaves the defaults. Panels keep four lanes across three providers, so diversity is unchanged. Efforts stay at high, the value #9's live probes used; raise per family in setup if wanted.

Open to changing any row before implementation.

Scope

• references/provider-dispatch.md: fold the three rows into the stock matrix; remove the "Additional model matrix" section; keep the route note (native spawn_agent under Codex, external runner under Claude Code).
• skills/setup-pstack/SKILL.md: new first-run sheet; drop "additional" wording; probe table rows become plain matrix rows.
• Skill and playbook text that names the old defaults (architect, arena, interrogate, poteto-mode SKILL.md and the feature / bug-fix / perf-issue / hillclimb playbooks).
• Tests: runner/model-matrix.test.ts (stock rows, first-run panel alignment, the "without changing the stock matrix" block) and the solo code roles stay on the sol row invariant in tests/skill-collision-repro.sh.
• Docs: README (rewrite as pstack-flex, one family table, no optional tier, no roadmap section), docs/LANES.md, docs/USAGE.md, docs/reference.md, UPSTREAM.md exclusions note, UPSTREAM-FLEX.md (the "stock matrix and first-run sheet are byte-unchanged" promise no longer holds; record the new conflict surface), CHANGES.md.

Consequences to accept

• Every upstream sync now conflicts on the matrix and first-run sheet. Cursor is moving its defaults to Opus 5.5 / Grok 4.7, not GPT-6, so these conflicts are permanent and resolved by hand each time.
• Existing sheets are untouched. Only a first run, or an explicit role change in setup, picks up the new defaults.
• Merge gate per AGENTS.md: the exact candidate installed in both harnesses, /setup-pstack first run producing the new sheet with passing probes for Astra, GPT-6 Sol, and Luna, and one mixed-panel smoke. Runner probes alone do not satisfy it.