ci: make the DOOP dataset manifest depend on the data, not the path

[justinjoy]

Refs #1297. Stacked on #1298 (fix/1294-manifest-collation) — both change the same line, so basing on main would guarantee a conflict. Merge #1298 first.

The defect

run-doop-perf-gate.sh hashed sha256sum "$data_dir"/*.facts. sha256sum prints <hash> <path>, and nothing stripped the directory — so one directory gave three hashes:

relative    871d21c0…
absolute    2887c99e…
trailing /  01e931b8…

Reported as DOOP dataset manifest … != pinned …, which reads as dataset substitution — the same false accusation #1294 removed from a different cause, in the same message.

The fix, and why the obvious one wasn't enough

( CDPATH= cd -- "$1" && sha256sum -- *.facts | LC_ALL=C sort -k2 | sha256sum | awk '{print $1}' )

Stripping the prefix afterwards was my first fix. GNU sha256sum escapes a path containing a backslash by prefixing the whole line, which a prefix strip can't remove — so two spellings of a backslash-containing directory still disagreed.

CDPATH= is load-bearing. cd consults CDPATH for a relative argument — which $data_dir is by default — and on a hit it prints the resolved path into the command substitution and reaches a different directory. With CDPATH pointing at anything containing a doop, the function returned that directory's hash and exited 0: a manifest whose purpose is proving the benchmark ran on pinned data, certifying data it never read.

-- covers option-shaped names; it does not cover a bare -, which bash converts to $OLDPWD after option parsing. The comment says so rather than claiming otherwise.

Compatibility

This changes the computed value. Nothing in-repo pins WIRELOG_DOOP_DATASET_MANIFEST_SHA256 — it's read from the environment and the gate skips when unset — so there's nothing here to recompute. But an operator holding a value recorded under the old formula sees a mismatch on the first run. The failure message now says the pin is stale rather than the data wrong, and names where the correct value already lives (downstream-matrix-oracles.tsv's files: field). CHANGELOG carries the same.

manifest_for_doop deliberately untouched

Its output is pinned in downstream-matrix-oracles.tsv; moving it inside a change premised on manifests not moving would be self-contradictory. The two agree on a canonical path and diverge on trailing slashes, dotfiles and symlinks — the self-test asserts the agreement and pins the divergences, so unifying them has to be deliberate. That unification is the shared helper the issue asks for.

Three defects in my own test, all one cause

defect	why it passed
backslash in the leaf	both spellings carried it, sha256sum escaped both identically — reverting to the broken implementation passed the assertion named for that bug
-x.facts in a two-spelling fixture	dropping -- broke both sides; two empty strings compared equal
'$data' single-quote interpolation	a quote in TMPDIR killed both subshells; two empty strings compared equal

Each was built from the shape of the description rather than the mechanism of the defect. Each is now built from the mechanism and killed by mutation.

Validation

• 11 assertions, linux-gated (the collation fixture can't discriminate where en_US is codepoint order)
• meson test -C build: 306 Ok / 0 Fail / 12 Skipped
• Mutations killed: drop CDPATH= (2 named assertions), drop sha256sum --, revert to the prefix-strip form, drop LC_ALL=C, name-dropping. manifest_for_dir correctly survives; cd -- survives and is documented as out of scope.
• Hostile environments all exit 0: nine TMPDIR shapes (space, quote, #, *, [, \, $, dot, relative), CDPATH=., CDPATH=/tmp, CDPATH=.:/tmp:/var, OLDPWD=/etc

Also fixes the SIGPIPE twin in check-manifest-collation.sh: locale -a | grep -Fxq reported the locale absent because grep exits first, and that gate's skip is exit 0 — so on a locales-all host it reported OK having asserted nothing.

Four review rounds; final verdicts Reviewer and Critic both approving.