refactor(rules): five worktree bullets each re-stated the same silent-failure frame (#193)

[ZacxDev]

Closes clawgate Task #193.

scripts/tests/test_rules_size.py stated in its own comments that eviction had been run to completion, and that the answer to the next real rule was consolidation, not another ceiling bump — naming "the four scattered worktree bullets" as the obvious candidate. PR #447 then added two recovered rules, taking the cluster to five. This does the consolidation it asked for.

What was actually duplicated

stated more than once	where	now
the git worktree add … -b <branch> origin/<main-branch> recipe	sync-a-branch bullet and the cross-repo bullet	once, in the git -C <repo> form that covers both; cross-repo references it
"never two file-modifying agents in one checkout"	as a hazard and as a prohibition, inside one bullet	once as each
the silent-failure framing ("no error, the worktree is created successfully")	spelled out by the cross-repo, session and .envrc bullets	hoisted once

Which bullets merged into which

Three separate "a worktree does not isolate X" rules became one frame with three enumerated surfaces:

🔴 But a worktree isolates a working DIRECTORY only — three surfaces it does NOT isolate, each failing SILENTLY (worktree created, no error):

• 🔴 the REPO it is built from … → archive: cross-repo-worktree
• 🔴 the SESSION … → archive: worktree-not-session
• 🔴 the ENVIRONMENT … → archive: worktree-envrc

Each surface keeps its own 🔴, its own imperative, its own failure shape and its own → archive: pointer — so none depends on the frame to carry its scope, and it survives being quoted in isolation. A fourth surface can join the list without re-stating the frame.

Left structurally alone: the sync-a-branch bullet (now holding the canonical recipe), the take-another-ref bullet, the base-clone re-sync bullet, and the pgrep/pkill bullet in Shell & Tooling — its /proc/<pid>/cwd clause is the one PR #447's audit had to restore, and it is untouched here.

A correctness fix, not only a byte one

The old text asserted "A worktree isolates the REPO, not the SESSION" while the bullet directly beneath it said the worktree is built from the wrong repo. Both cannot be true. "isolates a working DIRECTORY only" is the archive's own wording and is wider than either.

Byte delta

claude/RULES.md 33,300 → 33,229 B (−71). Headroom 1,900 → 1,971 (floor 900).

That is a small number and it is the finding, not a shortfall: every clause that survived is scope-bearing, so the only bytes available were genuine restatement. The frame itself costs ~78 B; it is paid for by ~150 B of real dedup. Recorded in the gate module so nobody reads this as evidence that another −71 B is sitting there — it is not, and the next real rule still costs ceiling.

What the 71 B bought: a fourth worktree hazard now attaches to the enumerated list for the cost of its own text, instead of arriving as a sixth sibling bullet re-stating the frame (~250 B of restatement, measured against the four it would have copied).

MAX_BYTES deliberately NOT ratcheted down. The slack predates this change; tightening the gate is a policy call that belongs in its own commit where it can be argued on its merits.

Nothing moved to the archive

There was no un-evicted narrative left in the cluster to move — the same "eviction has been run to completion" the module already recorded. The three archive sections had their *Supports:* quote lines updated to match the new core wording, and test_rules_size.py's comment quoting the now-superseded rule name was corrected (a comment is a claim too).

Qualifiers considered for trimming and deliberately KEPT

Each would have let the shortened rule pass while the hazard survived in a shape the original covered — the exact failure mode test_rules_size.py:128-133 forbids and that the #447 trim hit.

• "Two failure modes, only one loud" + both modes. The point of the pair is that you would not necessarily notice. Keeping only the silent mode loses the recognition tell; keeping only the loud one loses the dangerous case. Neither alone carries it.
• "that still satisfies the isolation mandate, so this is not an exception to it." Without it, "For cross-repo work do NOT pass isolation: "worktree"" reads as licence to skip isolation entirely.
• "file-modifying" and "in one checkout" on the never-two prohibition. A bare "never two" was written and then reverted — it silently dropped both qualifiers and left the antecedent to the reader.
• "Read-only agents (audits/research/exploration) don't need one." The parenthetical is what makes "read-only" operational rather than a judgement call.
• "working in a tree where its brief makes no sense." Arguably subsumed by the two failure modes — which is precisely the "it's covered elsewhere" reasoning the docs(rules): two multi-agent lessons that never landed, recovered from a merged branch #447 trim used, so it stays.
• The ../<repo>-<topic> path convention. Dropped when the recipe was unified into the -C form, restored once a cold read caught that it was concrete guidance and not framing.
• The per-surface 🔴 markers. Hoisting one marker onto the frame would have saved ~15 B and left each surface without its priority marker if ever extracted.

Verification

• All 5 functions in scripts/tests/test_rules_size.py pass (called directly — python3 on PATH has no pytest).
• Each negative-controlled, each going red with its own message: ceiling forced below current size (OVER its hard ceiling), headroom forced above the gap (no working headroom left), RULES_MD → missing file, ARCHIVE_MD → missing file, and a bogus → archive: no-such-anchor-here injected.
• A claim-coverage checker asserts all 56 distinct claims from the original 6-bullet block are still present, and records the 7 deliberate rewordings with the reason each is not a narrowing. It was itself controlled: empty file → all 56 reported missing (positive control); five single-claim deletions → red (negative controls); unmutated → green.
• Every → archive: anchor still resolves (test_every_archive_pointer_resolves).
• Authoritative gate nix build .#checks.x86_64-linux.pytests, read by content: TOTAL collected=9429 passed=9428 skipped=1 failed=0, all 17 targets above floor, no === FAILURES ===, no timeout panic, RESULT: PASS (exit=0).
• Cold read of every consolidated bullet as a reader who never opens the archive. This is what caught four narrowings, all fixed before commit: the dropped ../<repo>-<topic> convention, the bare "never two", "the flag" without an antecedent inside the sub-bullet, and the lost bold on the .envrc consequence clause.

🤖 Generated with Claude Code