fix(conformance): no hardcoded checkout path, and no silent substitute

[myobie]

This is how docs/conformance.md went stale, so it is not tidying.

Three sites defaulted to an absolute path on one developer's machine when PTY_NODE_CHECKOUT was unset. The path is a machine specific CLAUDE.md forbids in a shipped artifact — but the silence is what did the damage.

What it actually caused

The test did not skip. It compared against whatever was at that path, which on one host is a checkout seven commits behind, and reported a confident failure about a completion entry that was not missing at all. CI never saw it: the path does not exist on a runner, so there the filter returns None and the test skips. It only misbehaved where that path happened to exist — which is why it survived.

The generator was worse. With no checkout it fell back to that path, and node_suites falls back again to the compiled-in SUITES list, so running it with the variable unset wrote a checked-in document describing a Node tree nobody named, looking exactly as authoritative as a correct one.

That is the mechanism by which docs/conformance.md came to say 0.12.0+500eab2 while node-ref said 86dcc5e. A stale map is what a stale comparison is built on, and a stale comparison is what produced eight false divergences earlier this week.

Two cases, two different answers, on purpose

	behaviour when the reference is missing	why
a test	skips, naming the variable	the comparison cannot be made; that is not a failure
a generator	errors	it writes a file that outlives the run and carries no memory of what it was made from

completions.rs was also the only test doing this. parity_shapes.rs and parity_fixtures.rs already consult node_checkout() and skip. So this was a divergence from an established local convention rather than a missing one, and the fix copies the neighbours rather than inventing a pattern.

Controlled in both directions

A test that always skips would be worse than one comparing against a stale checkout — at least the stale one fails sometimes. So the skip path being obviously right is not a reason to skip the check:

• variable set to the pinned checkout: the test RUNS and passes, 9 of 9.
• variable unset: skips, printing no Node checkout with completions/ (set PTY_NODE_CHECKOUT).
• generator unset: exits 1, naming the variable and node-ref.
• generator pointed at a path with no tests/: exits 1 with that reason.
• generator pointed at the pinned checkout: gets past the checkout check.
• grep for /home/ and /Users/ outside target: 0 remaining.

One thing this surfaces but does not fix

Pointed at the pinned checkout, the generator now stops at a pre-existing guard: Node suite kill-report.test.ts has no classification. Five Node suites postdate the map — kill-report, output-activity, proc-table, process-groups, spawn-already-published. The SUITES table is untouched by this diff; classifying them is the separate work of regenerating the map, and it is exactly what the old silent fallback was hiding.