docs: close the knowledge-graph follow-ups

[wboayue]

Summary

Closes the last four follow-ups in plans/claude-md-knowledge-graph.md. Docs and one new rule node; no code.

1. Both live counts re-run, and both had moved

Claim	Was	Is
impl Client methods with # Examples	132 of 152	143 of 154, zero genuine gaps (#751)
<dir>/tests.rs residue	"about two dozen"	21, against 85 flat *_tests.rs

docs/rules/testing/sibling-test-files.md now carries the two find commands beside the numbers.

The durable finding is narrower than "counts rot". Across every count this arc corrected — 40 vs 24 text-framing tests, 42 vs 44 on_none sites, 49 vs 54 one-shots, 132/152 vs 143/154 examples, four vs 26 unnarrowed sites — the list was right and the total was stale in all but one. Lists are written by looking; totals are written by remembering.

2. One convention was living only in the memory store

A fluent builder's named methods must cover every variant of the enum they wrap — established by #549, with a bug class and a canary (_ => unreachable!() at a caller), and no node. It has one now: docs/rules/style/builder-enum-coverage.md, plus its index line.

Everything else in that store that reads like a convention was already covered by an existing node (branch-then-PR, toolchain pin, builder entry-point form, release-note attribution).

3. Memory-store reconciliation (outside this repo)

The drift was structural: 82 of 84 memories carried a name: that was a title rather than the filename slug wikilinks resolve against — which is exactly why 12 of 27 distinct links dangled, each a kebab spelling of an existing snake_case file. Names normalized, 19 links repointed, 21 backticked cross-references converted; zero dangling now. Recorded in the plan since the follow-up lived there.

4. Cluster subagents: decided, no

The precondition held (cluster boundaries survived all six passes), but the premise didn't. The case was stronger enforcement — and nothing in this arc was caught by a rule being loaded at the right moment. The routing gap (#730), the mispairing (#738/#749), the drifted TickDecoder consts (#739), the shadow submit_all (#750), the async retry inversion (#744) were each caught by a test that could fail, and each had been described accurately in prose for weeks beforehand. The graph carries why; the gates enforce. Revisit only if a cluster stops fitting in a read.

Also

CLAUDE.md's one-shot index line still advertised the hand-listed roster #749 deleted; corrected. Status header recounted against wc -w (1,625 words, 68% below the 5,012 the migration started from).

Verification

• just rules-check — 33 nodes, all links resolve
• cargo test --doc --all-features — 321 green
• cargo clippy --all-targets --all-features — clean