AGENTS.md: trim the sections that repeat other files, to cut every session's starting context #167

Problem

AGENTS.md loads into every agent session in this repository: 22,230 characters, about 6k tokens. In a fresh session it is most of the roughly 10k tokens of memory files, out of a starting context of about 42k (measured 2026-10-29 with two Sonnet sessions and this session's context breakdown). Several sections repeat what another file already holds, or detail what an agent only needs at one moment:

• Finding code (~920 tokens): most of it is the dated list of what the code graph misses, as measured on 2026-10-13.
• Commands (~920): the BuildVerdict, BashGuard and session-start paragraph repeats what the guard prints when it refuses a command, and decisions 92 and 105.
• Issues (~450): largely CONTRIBUTING's "Issues and labels" and "Pull requests", plus the maintainer's global tracking rule, which also loads into every session.
• Workflow step 7 (~400): the oracle-kit mechanics that test/oracles/README.md already holds.
• Releasing (~300): CONTRIBUTING's "Releasing" step 1.

Change

Trim those five sections to what an agent must have before it acts, and point to where the rest lives. Move the graph measurements into a design note of their own, linked from design/plan.md. Keep "Rules that are easy to break" word for word. Keep every heading another document cites: decision 93 cites "Finding code".

Done when

• AGENTS.md is shorter, and nothing it drops is missing from the file it points to. Measured: 22,230 to 17,956 characters, about 1.1k tokens, not the 2k first estimated (docs: AGENTS.md keeps what an agent needs before it acts, and points to the rest #168's table has the sections).
• The conventions tests and docs-query check pass.