Avoid needless completeness queries and cap CI database concurrency

[roed-math]

Summary

• Add an optional database-independent precheck phase to
  CompletenessChecker.
• Use it to recognize incompatible number-field root-discriminant and Galois
  root-discriminant constraints before querying for null stored values.
• Add regression coverage proving that the contradictory query does not access
  the database and that compatible intervals are not short-circuited.
• Cap matrix execution at six concurrent jobs to reduce peak load on the shared
  devmirror database.

Motivation

With the statement timeout introduced in #7181, UtilsTest.test_complete
repeatedly failed while executing this preliminary null-data query
(e.g. https://github.com/LMFDB/lmfdb/actions/runs/31268853079/job/93204410081):

SELECT "id"
FROM "nf_fields"
WHERE "degree" = 5
  AND "rd" >= 40
  AND "rd" <= 60
  AND "grd" IS NULL
LIMIT 1;

The original query is mathematically impossible: its root-discriminant interval
lies strictly above its Galois-root-discriminant interval. The existing
NFBound logic detects that contradiction, but only after the generic null-data
probe.

This PR introduces a narrow precheck contract for conclusions that are valid
independently of database contents. It does not move ordinary completeness
guarantees ahead of null checking: those can depend on all relevant columns
having been computed, so they stay behind the null-data check.

Safety

• Predicates carrying SQL-null semantics (None anywhere in the rd/grd
  value, e.g. {"$ne": None} for IS NOT NULL) bypass the numeric precheck
  entirely: the real-number model cannot represent them (to_rset reads
  None as the whole real line, so {"$ne": None} would collapse to the
  empty set and misreport a nonempty search as impossible). A sentinel keeps
  an omitted constraint distinguishable from an explicit null predicate, and
  a regression test covers four null-predicate shapes, including the
  {"$or": [{"$ne": None}, {"$lte": 30}]} collapse.
• Compatible queries still use the existing null-data checks.
• The existing fallback to "completeness unknown" is unchanged.
• The test expectation is not weakened.
• The workflow's 60-minute job timeout is unchanged.
• No database index or schema change is included.

Verification

• Focused no-database-access, compatible-input, and null-predicate regression
  tests pass (22 tests total in test_utils.py).
• Calling results_complete on the contradictory query against the real
  database with psycodict statement logging enabled executes zero SQL
  statements and returns in under a millisecond, with the exact existing
  result triple.
• tox -e tests -- lmfdb/tests/test_utils.py passes (21 tests, 19.7s), with
  UtilsTest.test_complete at 14.9s.
• The same targeted test passes with PGOPTIONS='-c statement_timeout=120s'
  (the pre-Raise the CI database statement timeout to 10 minutes #7184 threshold) and with PGOPTIONS='-c statement_timeout=10min'.
• Workflow YAML parses; jobs.test.strategy retains fail-fast: false and the
  unchanged matrix expression, with max-parallel: 6 added and
  timeout-minutes: 60 untouched.
• git diff --check, pyflakes, pylint (repo flags) and
  ruff check --preview --select=E722 pass on the edited files.
• Full CI: all 12 devmirror matrix jobs passed on both the original commits
  (run 31300780554,
  16 minutes total) and the null-predicate fix
  (run 31332813750).
  Details below are from the first run:
  • The formerly failing shard (devmirror belyi higher_genus_w_automorphisms hypergm lattice maass_forms tests users, the one running lmfdb/tests/test_utils.py) passed.
  • UtilsTest.test_complete took 9.22s in CI (it previously stalled on the canceled null probe until the 120s statement timeout); the two new regression tests took 0.00s.
  • No statement timeout or QueryCanceled appears anywhere in that job's log.
  • All matrix entries were still created; max-parallel limited execution without suppressing jobs (observed 6 in progress + 6 queued, with the second wave starting as first-wave jobs finished).

🤖 Generated with Claude Code

[roed-math]

Pushed 2e6a262 fixing a correctness issue in the precheck: psycodict predicates carrying SQL-null semantics ({"$ne": None} = IS NOT NULL) were misread by the numeric model as empty sets (to_rset maps None to the whole real line, so its complement is empty), so e.g. {"grd": {"$ne": None}} — and the nastier collapse {"grd": {"$or": [{"$ne": None}, {"$lte": 30}]}} — could be declared intrinsically impossible without touching the database.

The precheck now conservatively declines any rd/grd constraint containing None anywhere in its expression tree, using a sentinel so an omitted key (no constraint) stays distinguishable from an explicit null predicate. The numeric contradiction fast path is unchanged (still zero SQL statements for the motivating query), and a table-driven regression test covers the four null-predicate shapes. NumberSet/to_rset themselves are untouched.