Add isolated agent experiment layer

[tim-codes]

Experimental, vendor-neutral agent context layer: a portable policy/capability contract, profiles (exxo, exxo-h2gc-portal), a codex-air adapter, and scripts/agent-experiment (inspect / validate / render / stage). Render-only by design: it never writes to live Claude/Codex/Air config or downstream repos. stage writes only to an explicit empty scratch directory.

Findings are in docs/agent-experiment-findings.md. In short, go for continued render-only experiments and no-go for live Codex/Air installation until Air's instruction/skill discovery is probed.

Kept as a draft, not for merge. Open points:

• tests/agent-experiment.bats has not been run (bats isn't installed); it would be the repo's first test suite.
• The diff contains local absolute paths and Exxo client repo references.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Nk5birdBvHk51jTgSp8oL3

[tim-codes]

Parked: not merging for now. Reviewed 2026-10-11 alongside #107/#108.

Blocking issue: tim-codes/dotfiles is a public repo, and this diff carries

• Exxo client references: the exxo-h2gc-portal profile, h2gc instruction and skill pointers, and the findings doc discussing the h2gc repo
• local absolute paths (/Users/tim/...: ~/.codex, Air.app and its Application Support/Caches/Logs, ~/dev/exxo/... skill sources)

Because the branch is pushed, that content is already publicly visible, not just on merge.

Other open points (from the description, still true): tests/agent-experiment.bats has never run (bats isn't installed; it would be the repo's first test suite), and the branch is ~57 commits behind main.

Decision: leave this draft as is. Come back later to rethink where client-specific and path-bearing material should live (a private Exxo repo, or a scrubbed generic version here) and how to clean the git history/branch, before any merge.