Agents: require style and length consistency with upstream

[moshap-firebolt]

AGENTS.md currently asks for succinctness only in the PR comment. Everything else an agent writes — comments, commit bodies, test tables — is unconstrained, and LLM agents default to far more prose than this project uses. That matters here specifically because this fork's changes are expected to go upstream, where verbose contributions get rejected.

Adds a Matching Upstream Conventions section with concrete budgets taken from this repository's own history rather than adjectives, so an agent can check its own output:

• a typical upstream parser fix adds 5–7 comment lines
• most upstream commits have no body; the longest run to ~20 lines
• no Firebolt ticket IDs in code or messages, since they mean nothing upstream

Plus a short Reviewing Agents subsection — arguably the higher-leverage half, since a reviewing agent's output lands directly in a PR thread.

Worth saying where this came from: the two parser PRs I opened this week both overshot. #10's original commit body was 30 lines, longer than any of the last 60 commits in this repo, and its PR description was several times the length of the surrounding ones. I've since retitled and trimmed it. The guidance is what I'd have wanted to read first.

---

Note

Low Risk
Documentation-only change to AGENTS.md; no code, security, or runtime impact.

Overview
Extends AGENTS.md so agent output (comments, commits, PRs, tests, reviews) matches this fork’s terse upstream style, not default LLM verbosity.

Adds Matching Upstream Conventions with concrete budgets: short why comments, dialect-style commit subjects with little or no body, no Firebolt ticket IDs, and tests that pin behaviour without extra tables or commentary.

Adds Reviewing Agents: few sentences per real issue, classify bug vs behaviour change vs suggestion, and skip fmt/clippy nits.

^{Reviewed by Cursor Bugbot for commit 12e8f2e. Bugbot is set up for automated code reviews on this repo. Configure here.}

[moshap-firebolt]

No issues. The budgets are checkable, which is what was missing.

Suggestion. #10 still overshoots them (the 9-line block on parse_in, the test preambles, measurements in the commit/PR). Trim that PR to match this file once this lands — otherwise the first thing an agent reads is a counterexample.