Route agent instructions instead of restating them

[tony]

AGENTS.md is read on every task, and CLAUDE.md symlinks to it, so all 364 lines loaded whatever was being changed — the MCP swap gotchas and the comment policy included. A root instruction file should answer which policy applies to the thing being changed, not hold every policy itself.

This splits the policy in three and leaves the root file as a router.

Layout

• AGENTS.md — a short router: change discipline, then one route per class of change, and nothing else.
• .github/WRITING.md — how we write: voice, README, changelog, release notes, commit messages, XML documentation, source comments, terminology, Markdown, code blocks, error messages, slop prevention.
• .github/CONTRIBUTING.md — how we work: setup, the socket root, building, tests, the gates, the MCP server, what a change must carry, pull requests, review, releases, compatibility.

The root CONTRIBUTING.md is gone; .github/CONTRIBUTING.md replaces it. GitHub reads a contributing guide from that location, so the new-issue and new-pull-request pages are unaffected. CLAUDE.md follows the symlink with no separate change.

Routes are phrased as conditions — "for changes to X, follow Y" rather than a bare "read Y" — so an agent re-evaluates per change instead of treating the pointer as a one-time startup instruction. Each listed file is stated to be the single home for its subject, so a rule that appears twice has one governing copy.

This is the content pass, not a skeleton

An earlier version of this branch landed the layout with every section body as the literal string <placeholder>. That commit is gone; the branch is rebuilt from master as six commits carrying real content. Nothing in the tree says <placeholder> any more.

The prose is drawn from what AGENTS.md and the old CONTRIBUTING.md already said, plus what the tree actually does. Where the two disagreed, the tree won:

• The toolchain claim was wrong. AGENTS.md said dotnet "resolves through mise". There is no mise.toml in the repo, and CI does not use mise at all — it installs the pinned SDK with actions/setup-dotnet from global-json-file. mise exec -- is a local convention, and CONTRIBUTING.md now says so.
• Two validators were documented; there are five. verify_workflows.py, sync_snippets.py --check and dump_tools.py --check were not mentioned anywhere.
• <example> is not a house convention. It appears exactly once in 27k lines of src/. WRITING.md says so rather than recommending it.
• How tested examples work was undocumented. The only account of it was a docstring in sync_snippets.py and an XML comment on a test class. examples/README.md is new and covers it where somebody editing an example already is.
• The ≤50-character subject rule is not followed. 24 of the last 60 subjects exceed it, the longest at 75. WRITING.md records ≤72 with ≤50 preferred, matching practice. No history is rewritten.

Three fixes so the guides describe the repository

• CHANGELOG.md — alpha.5, alpha.6 and alpha.7 were bracketed like links with no reference definition, so they rendered as literal text. All three tags exist; the definitions are added.
• README.md — titled libtmux for .NET, matching every sibling port, and the alpha warning now states the terms the family shares: an -alpha prerelease tag, an unsettled API, no deprecation period, pin an exact version, not recommended for production. The previous wording also claimed behaviour is proven "on every commit", which the tmux matrix does not do — it runs on its own workflow.
• Install verb — dotnet add package becomes dotnet package add in the root README, each package README, and the quality bar that quotes it. Verified against the pinned SDK (10.0.302), which accepts the noun-first form and its --prerelease flag.

How examples are verified

Two mechanisms, catching different failures, now written down in examples/README.md:

• ReadmeExampleTests compiles every C# block in the nine shipped documents against the real assemblies, and executes the ones tagged ```csharp run against a tmux server of their own.
• sync_snippets.py + SnippetContractTests cover drift: a block anchored with <!-- snippet: Name --> is materialized from a #region of the same name, --check fails the build on any hand edit, and every published region must be an [Example] method that runs.

CONTRIBUTING.md lists the sync step beside the check, and WRITING.md says which tag to use and why.

Verified

• dotnet build LibTmux.slnx --configuration Release --warnaserror — 0 warnings, 0 errors.
• sync_snippets.py --check — 9 snippets current.
• verify_public_api.py, verify_capabilities.py, verify_workflows.py — all clean.
• pytest eng — 10 failed, 1103 passed. Identical on master, confirmed by running the same suite at origin/master in a scratch worktree. The failures need the Python libtmux checkout that CI clones and a local run does not have; they are not from this branch.
• No <placeholder> anywhere; every relative link in the three documents resolves; CLAUDE.md still resolves to the router; no line over 80 columns; no GitHub alert blocks.

Deliberately not done

• Anchoring more blocks to snippet regions. Only one block in the root README is anchored today. Everything else is still compiled and run, so nothing is unverified — anchoring would only add drift protection, and is worth doing per block rather than in bulk.
• The other six ports. Rust, Java, TypeScript, Swift and C++ still hold the placeholder skeleton, and the family is split on alpha-warning markup — TypeScript, Swift and C++ use > [!WARNING] blocks that .NET, Rust and Java do not. Untouched here.
• State the comment bar in CONTRIBUTING #9 edits the root CONTRIBUTING.md this branch deletes. Its prose is a good candidate for the source-comments section; merge order is a judgement call.