Stage 3 takes the live years from tmean_annual alone #121

Found by the plan review for #119.

What changes if we do it

scripts/pipeline_stage3_edh.R checks its publish against the years every live
COG holds, the way pipeline_update_edh.R STEP 1 does since #119, so its result
does not depend on which COG a dead sync happened to upload.

What happens if we never do it

Stage 3 takes required_years from the live tmean_annual COG alone
(pipeline_stage3_edh.R, the live_years <- tryCatch(...) block before
publish_problems()). After a STEP 5 sync that died partway:

• tmean_annual ahead (holds N, others N-1): stage 3 refuses unless the local
  backfill holds N. That is safe, but the refusal reads as a backfill problem, not
  a partial sync.
• tmean_annual lagging: stage 3 publishes 1950 to N-1 over COGs that hold N. That
  is benign, since N was never in the catalog, but the year is silently dropped
  from the COGs that had it.

pipeline_update_edh.R now repairs a partial sync on its own (#119), so stage 3
is only reached by hand. Low priority.

Suggested fix

Read all 59 live COGs' band names as pipeline_update_edh.R STEP 1 does, and pass
sort(unique(unlist(...))) as required_years, using live_spans() in
scripts/_lib.R.