feat(pstack): select requested effort per model family

[ericlitman]

Closes #15.

What changed

• defines selectable and default effort levels for Fable, Sol, Grok, and Opus in one provider model matrix
• makes setup-pstack rerunnable for per-family effort changes while preserving role assignments
• validates and live-probes all four exact provider/model/effort pairs before any write
• rejects mixed, invalid, duplicate, or unknown configuration state instead of choosing precedence
• mirrors the Codex sheet inside one bounded AGENTS.md block and keeps Claude's single include
• adds native Claude agent definitions for every Fable and Opus effort level
• bumps the plugin to 1.1.0 and updates reference and release documentation

Verification

• bun install --frozen-lockfile
• bun run test: 99 passing, 445 assertions
• strict TypeScript typecheck
• PSTACK_STATIC_ONLY=1 bash tests/skill-collision-repro.sh
• strict Claude plugin, marketplace, Codex plugin, and agent validation
• exact source-to-install byte parity in Claude Code and Codex
• Claude Code: all 10 native Fable/Opus agent variants plus six setup transitions and a mixed-panel cross-judge
• Codex: six setup transitions, 22 complete external receipts, native Sol probes, exact one-block AGENTS mirroring, and a mixed-panel cross-judge
• unchanged reruns were byte-identical; single-family transitions changed only that family's descriptors

No fallback or implicit timeout was used. Live evidence proves requested route and effort flags, not hidden provider reasoning depth.

[mergify]

Tick the box to add this pull request to the merge queue (same as @mergifyio queue).

• Queue this pull request

[ericlitman]

Independent shipping verifier: PASS

Verified head: e6ed855.

I checked the Issue 15 body and canonical triage acceptance criteria against the diff and the exact installed 1.1.0 candidate.

• Repository gates pass independently: 99 Bun tests with 445 assertions, strict TypeScript typecheck, static invariants, manifest parsing, strict Claude plugin and agent validation, and git diff checks.
• The source plugin matches both the Claude Code and Codex installed 1.1.0 trees byte for byte, excluding only dependency directories and the Claude runtime .in_use marker.
• A fresh read-only Claude Code session loaded the exact installed 1.1.0 plugin and resolved all ten expected Fable and Opus native agent types. Every exact type returned its unique marker.
• The Codex evidence recalculates to six passing setup transactions, 24 independent effort questions, 22 complete external receipts, seven native Sol probes, exact one-block AGENTS mirroring, and all provider probes completing before each write. The unchanged rerun is byte-identical. Fable, Sol, Grok, and Opus transitions each changed every occurrence of only the selected family while preserving role and lane coordinates.
• The Claude evidence shows the same six-run default, unchanged, Fable, Sol, Grok, and Opus sequence with exact readbacks, one stable CLAUDE include, family-only diffs, and a passing final mixed panel plus cross-judge.
• The setup contract loads current state before selection, rejects duplicate, unknown, unmatched, out-of-domain, and mixed-effort state before probes or writes, probes exactly four selected pairs, snapshots and rolls back both artifacts, preserves role families, and uses the declared matrix as its sole effort source. Runner and receipt evidence passes requested effort exactly with no fallback or implicit timeout and does not claim hidden applied reasoning depth.

No acceptance blocker found.