chore(deps): bump nixpkgs + home-manager to 26.05-darwin

[JacobPEvans-personal]

Summary

Bumps the NixOS channel from 25.11 to 26.05 across the flake, keeping nixpkgs and home-manager in lockstep, and resolves a home-manager 26.05 breaking change that collided with this repo's legacy skills option aliases.

Both channels now on 26.05:

• nixpkgs → nixpkgs-26.05-darwin (top-level flake.nix + modules/mcp/flake.nix)
• home-manager → release-26.05 (release branch must match the nixpkgs channel)
• lib/versions.nix stableVersion → 26.05 (informational, consumed by nixos-release-check.yml / ci-eol-check.yml)

home-manager 26.05 breaking change + migration

home-manager 26.05 introduced a native programs.codex.skills leaf option (modules/programs/codex.nix), typed either (attrsOf (either lines path)) path, that manages codex-only skills under ~/.codex/skills. A leaf of that type "does not support nested options."

This repo's modules/agent-skills/default.nix carried legacy mkRenamedOptionModule aliases redirecting programs.codex.skills.{fromFlakeInputs,local} → programs.agentSkills.*. Those aliases declare child options under the new leaf, so the module system now throws:

error: The option `programs.codex.skills' ... would be a parent of the following
options, but its type ... does not support nested options.
 - option(s) with prefix `programs.codex.skills.fromFlakeInputs' in module `.../modules/agent-skills'
 - option(s) with prefix `programs.codex.skills.local'          in module `.../modules/agent-skills'

This failed the checks.x86_64-linux.module-eval CI check (and every other home-manager-driven regression check) on the Linux runner. Confirmed introduced by the home-manager bump: the same eval succeeds on origin/main (home-manager 25.11).

Migration chosen — drop the colliding codex aliases. nix-ai's real feature is programs.agentSkills.*, a cross-tool skill deployer targeting ~/.agents/skills — distinct from home-manager's codex-only programs.codex.skills. The aliases were pure backward-compat redirects; nothing in this repo (or in nix-darwin) sets the legacy programs.codex.skills.* paths, so removing them loses no configuration. The programs.gemini.skills.* aliases are kept — home-manager exposes gemini skills at programs.gemini-cli.skills, not programs.gemini.skills, so those do not collide (verified empirically). stateVersion test fixtures are intentionally left at 25.11 (plain state-compat markers, no channel-coupling guard — lagging is supported).

Verification

• nix eval .#checks.x86_64-linux.module-eval.drvPath → returns a valid .drv (previously threw).
• All home-manager-driven regression checks evaluate: defaults-regression, options-regression, and the {agent-skills,codex,gemini,mlx,fabric}-defaults-regression checks.
• nix flake check clean (native; cross-system check builds are skipped on darwin as designed — CI runs them on Linux).
• pre-commit (nixfmt, statix, deadnix) passes; no hooks bypassed.

FYI (not addressed here, out of scope)

home-manager 26.05 also renamed programs.codex.custom-instructions → programs.codex.context. modules/codex/default.nix:31 still uses the old name; home-manager emits a soft deprecation warning but evaluation succeeds via its own rename alias. Behavior preserved. Worth a follow-up to adopt programs.codex.context.

🤖 Generated with Claude Code