docs: the generated env-var table and its count line live in CLAUDE_RULES.md, so every env-var PR conflicts with every other #16317

Problem

Every PR that adds or removes an AUTOBOT_* env var rewrites the same line of docs/developer/CLAUDE_RULES.md. So any two of those PRs always conflict, not just sometimes.

Evidence

• Merge train train(2026-09-11): one CI run for 14 approved PRs — test vehicle, do not merge #16314 (2026-09-11) dropped 3 of 17 approved PRs on conflicts. All three conflicted in docs/developer/CLAUDE_RULES.md: feat(slm): nodes report their GPUs in the heartbeat, and the SLM serves them per node (#16280, #16281) #16292, fix(llc): kill an agent's whole process tree on cancel, timeout or stall (#13097, #13099) #16284 and feat(agents): renew work claims only while the run makes progress, and refuse writes under a lapsed claim (#15950) #16273.
• fix(llc): kill an agent's whole process tree on cancel, timeout or stall (#13097, #13099) #16284 and feat(agents): renew work claims only while the run makes progress, and refuse writes under a lapsed claim (#15950) #16273 edit the same line. Each rewrites *219 variables registered as of last generation.*, one to 221 and the other to 220.
• pipeline-scripts/generate_env_docs.py:69 writes that count note. pipeline-scripts/check_env_var_registry.py:204 regenerates it and requires an exact match, and that check runs in pre-commit and code-quality.yml. So the line must change on every env-var PR.
• The table's rows are sorted. Two inserts at neighbouring positions conflict as well.
• The table crowds out the rules. CLAUDE_RULES.md is the rules doc that agents load before starting a task (the trigger table in CLAUDE.md). 220 of its 669 lines are this generated env table.

Options

• A. Drop the count line. The checker already compares the table against the registry, so a count adds nothing the table doesn't show. This removes the always-conflict; neighbouring-row conflicts remain.
• B. Move the generated section to its own generated doc, leaving a one-line pointer in CLAUDE_RULES.md. The rules doc stops being a merge hotspot, and agents stop loading 220 rows along with the rules.
• C. Mark the generated file merge=union in .gitattributes. Not recommended: a union merge can duplicate or misorder rows. The checker would catch that loudly, but it trades a conflict for a red.

Recommendation: A + B.

Acceptance criteria

• No tracked line changes on every env-var add or remove: the count note is gone, or derived when read.
• The generated env table lives outside CLAUDE_RULES.md, and CLAUDE_RULES.md links to it.
• generate_env_docs.py and check_env_var_registry.py target the new file, and pre-commit and code-quality.yml are green.
• Every reference to the table's old location is updated, including the scripts' messages ("Stage the updated CLAUDE_RULES.md…") and the docs.
• Evidence: two branches that each add one env var at non-neighbouring sort positions merge in either order without conflict. Show it with git merge-tree.

Out of scope

The two file-size ceiling lists (scripts/python_file_size_known_large.py and repo_tests/python_file_size_ratchet_baseline.py), which #16284 also conflicted on, are a deliberate second copy. repo_tests/python_file_size_ratchet_baseline.py:10-19 explains why: "so no ceiling can be raised by editing one file alone". Their conflicts are neighbouring-line conflicts by design.

Sequencing

This lands after the three dropped PRs are rebased and merged, so they don't have to rebase twice.