test(corpus): exploration query family (q055-q059) — naive first-touch profiling

[aaj3f]

What

Adds the five naive first-touch queries every human or LLM agent runs on an unknown dataset — the BI-derived corpus had zero pure-wildcard coverage. New closed Tag variants Wildcard and Exploration; native oracles blessed; honest per-member status taken from the actual run.

id	query	native	virtual	rows	gate
q055	?s ?p ?o LIMIT 5	1 ms	2.44 s	5	rows_only
q056	COUNT(*){?s ?p ?o}	0 ms	168.29 s	1	full
q057	DISTINCT ?p LIMIT 100	0.9 s	34.93 s	100	rows_only
q058	?s a ?type census GROUP BY	0.15 s	5.40 s	16	full
q059	predicate histogram GROUP BY ?p	2.6 s	32.38 s	10	full

Measured virtual-sf01 vs native-sf01, hot cache, 1 rep, on the #1503 head (6750286df). All five complete ok — none DNF (the predicted DNFs for q056/q058/q059 were overturned by measurement); expected_status stays ok/ok for all five.

Mechanisms

All five hit the same whole-dataset breadth (~69,046 files across all 16 TriplesMaps — the maximum-fan-out first-touch case). The wall splits by what columns get materialized:

• q056 (168 s) is the headline gap — predicate-less COUNT(*) has no manifest COUNT shortcut (not a single resolved table), so it materializes every triple: full object-column fetch, fetch_bytes 63× that of q057 over the same file set. It reads object values it only needs to count — same family as q038's un-fused COUNT. A fused predicate-less COUNT(*) (Σ over TriplesMaps of rows × POMs from manifest record counts) would take it to sub-second (PR-6 family).
• q057 / q059 (32–35 s) — only ?p needed → object columns pruned → per-file footer/decode overhead across 69,046 files dominates. This is PR-2a (the per-file decode-wall) at maximum fan-out.
• q058 (5.4 s) — ?s a ?type needs only subject-key + per-table rr:class constant → 16 scans (one per TriplesMap), column-pruned. The cheapest wildcard, and why the census DNF prediction was wrong.
• q055 (2.4 s) — LIMIT 5 doesn't prune the scan but stops after one fact table (~7,679 files, 63 MB).

Gate decisions

• q055, q057 → rows_only. Both truncate an unordered set with LIMIT, so two correct engines return different-but-valid subsets that can't be hash-compared. q057 is empirical — it was the sole parity failure in compare, and the mapping declares ~120 distinct predicates (119 rr:predicate + rdf:type) > the LIMIT 100, so the 100-subset is nondeterministic.
• q056, q058, q059 → full hash; all three matched native exactly. q059 is deterministic despite the LIMIT (ORDER BY DESC(?n) ?p tiebreak).
• q056 timeout_s 300 — measured 168.3 s hot; 300 s is CI-stability headroom (168/180 = 93% is fragile), not an expectation. The wall is recorded honestly in the ROADMAP row.

Commits

1. test(corpus): fix error-boundary drift — declared_error must include q013 — a pre-existing test/manifest drift: the manifest declared q013 virtual=error but the error_boundary self-test's expected set listed only q034/q051. Isolated as its own commit.
2. test(corpus): add the exploration query family (q055-q059) — the two tags, 5 .rq files, manifest entries, and blessed native oracles.
3. docs(roadmap): exploration family section — measured walls + mechanisms, kept OUT of the north-star 44/51 accounting pending AJ's ruling on whether wildcard first-touch queries join the ≤3 s bar.

Stacked on #1503 (perf/r2rml-loadtable-meta-cache). Additive-only — no existing blessed oracle or manifest entry changed. Corpus self-tests green (7 passed).