Page the dark-streak history past cancelled runs

[ppXD]

Summary

• The real-model lane's consecutive-dark check read a single page of 12 completed main runs. After a burst of pushes most of that page is cancelled runs with no census, so it found too few census-bearing predecessors, read a short streak, and the lane went green while measuring nothing: main run 36565851194 concluded success with two UNMEASURED clauses in the middle of a 69-run dark streak.
• The run listing is now paged (page=1,2,…, 12 per page) until every unmeasured clause is settled, meaning its streak was cut by a run that measured it or already reached DARK_RUNS_TO_RED, or until DARK_HISTORY_RUNS_CAP=60 runs have been scanned (most dark streaks settle on page 1 or 2). Runs without a census are still stepped over and never count as a reset, evidence is counted per clause (a census-bearing run that never mentions the clause says nothing about it), and a run listed on two pages is read once.
• A scan that ends unsettled now warns that the check is INCONCLUSIVE and names how many census-bearing runs it found in how many it scanned, instead of reporting the short streak as the answer; the step summary's Streaks read from X of the Y predecessor runs scanned line carries the same paged numbers. DARK_RUNS_TO_RED, the census header, the branch rule and the summary table are unchanged.

Test plan

• bash .github/scripts/assert-every-filter-clause-ran.test.sh: 64 checks pass, the 43 existing ones unchanged plus 21 covering paging, early stop on a measuring run, the cap (runs, not pages), the INCONCLUSIVE warning, de-duplication across pages, per-clause evidence, and the pinned cap constant
• The new cases fail against the previous guard (the regression case exits 0 where it must exit 1)
• Mutations each fail the suite: single page, no cap (outer, inner, both), a measuring run no longer settling a clause, a count-based stop, no de-duplication, no INCONCLUSIVE warning, a red-length streak no longer settling, no scanned count in the no-census message, and removing the end-of-history break
• On the next real-model run on main, check that a dark clause's guard step reports its streak over the paged scan