study_area_run.sh: a run that stops before its tail leaves no record of which groups it was running #358

Problem

#348 lets a run that loses groups still recompute and compare the rest. It also adds --recompute-only --run-uid=<uid> to finish a run once the missing groups have been re-modelled. Two gaps remain.

• The expected set has no record of its own. Without --wsgs=, --recompute-only expects the groups that have a <schema>.log row for the run_uid on the dispatcher. Some groups have no such row:

  • a group that failed on a cloud host, since consolidate copies log rows only for groups that finished;
  • a group that failed before .lnk_log_run_start(), for example in lnk_load_overrides(), at the downstream guard's quit(1), or on a connect failure (wsg_run_one.R).

  The driver knows the full set once the buckets resolve, but it writes that set nowhere.

• A consolidate failure still exits before coverage. If one source is refused (dest_log_newer, which a single group can trigger), the driver exits after consolidate, before the coverage step and the recompute. That is the same failure class as One failed group skips the post-consolidate recompute for every group #348. Since One failed group skips the post-consolidate recompute for every group #348 the banner prints the --recompute-only recovery, but the groups that were consolidated are not recomputed in that run.

Proposal

• Write <TS>_run_wsgs.tsv (group, host, SKIP) as soon as the buckets resolve. --recompute-only reads it by default (<run_uid minus its suffix>_run_wsgs.tsv, since a run_uid's prefix is its TS) and falls back to the log.
• After a consolidate refusal, carry on to coverage with the sources that consolidated, keep the refused host, and exit non-zero as the kept-host path already does.

Acceptance

• --recompute-only with no --wsgs= names a group that failed on a cloud host.
• A run with one refused source recomputes the other hosts' groups and exits non-zero naming the refused host.

As built (PR #359)

• <TS>_run_wsgs.tsv is written before any spend and rewritten after per-host accounting with each host's SKIPs. --recompute-only reads the file together with the log, not instead of it.
• A refused or failed source no longer ends the run when every source has a recorded outcome (<TS>_consolidate.tsv). Its groups get the new coverage class not_consolidated and are kept out of the recompute.
• The refused host is kept only while finished groups remain on it. A host whose database holds nothing to lose is released and burned, as study_area_run.sh counts finished groups from log text and burns hosts holding unconsolidated work #347's rule says. The run still exits non-zero.
• A crash mid-copy is covered too. <TS>_consolidate_state.tsv is written in_flight before the copy, and a later bare --recompute-only keeps any group not ok out of the recompute until --wsgs= names it.