fix(ci): stop weekly-refresh dying at the 6h ceiling before it dumps

[Seungpyo1007]

What is broken

Every weekly-refresh run ends at 6:00:2x — the GitHub job ceiling — always inside Enrich benchmarks (all sources):

34113867834   6:00:23  cancelled
33389588839   6:00:21  cancelled
32697778694   6:00:21  cancelled
32001853491   6:00:21  cancelled
31364447360   6:00:22  cancelled
30799400591   6:00:23  cancelled

A job that hits the ceiling is reported as cancelled, not failed, so it never showed up in the failed-run list. Meanwhile every step after the scrape was skipped:

cancelled  Enrich benchmarks (all sources)
skipped    Validate (schema / range / slug / FK)
skipped    Integrity check
skipped    Verification band report (Tier 0)
skipped    Generate static dump
skipped    Create refresh PR

What it cost

The static dump has been drifting away from data/ for at least six weeks, because the only job that regenerates it never reached that step. Measured against the current develop:

collection	pages	verified disagrees with the record
gpu	2,030	565
pda	140	39
brand	207	24
soc	2,104	23
watch	433	7

The published API is advertising a verification state the data no longer has — 28% of GPU pages. (Promotion PRs flip verified in data/ and do not touch site/public/v1; the weekly dump is what was supposed to reconcile them.)

Change

• Collection gets a 200-minute budget, checked between sources, so validate / integrity / dump / PR always get the remainder. A source is never cut mid-scrape — the budget is only consulted before starting the next one.
• The source order rotates by ISO week. With a fixed order the tail (topcpu-cpu, topcpu-gpu) would be starved forever once the budget bites. Verified: week 37 starts at geekbench-nbc and wraps.
• timeout-minutes: 330 on the job, so the platform ceiling can never silently truncate it again — it fails loudly instead.

bash -n on the rendered step body passes; the YAML parses.

Refs #1