download/ingest housekeeping: close #5 and #2, fix #33, bench70 unit, old solar backup

[bob-ionis]

Part of the download/ingest/bronze plan (tracking issue linked below).

• close PLAN: backfill wspr.bronze 2026-06 → present (~660M spots) from wspr.live #5 (WSPR backfill 2026-06→present). Done: 2026-06..09 hold 200–218M rows/month
• close contest-ingest: add timestamp validation to reject typo'd dates #2 (reject typo'd dates at ingest). Superseded: bronze keeps them, tagged off-declared-year / timestamp-unrepresentable
• fix internal/wspr: go test fails at setup — bridge.c in a package that does not use cgo #33 so make test passes and can gate a release again
• clear the failed llm@bench70.service
• drop solar.bronze_pre_rebuild_20260922 once Judge accepts the solar rebuild

Validated when

• every box above is checked, and systemctl --failed is empty of lab units

[bob-ionis]

Tracking: #42

[bob-ionis]

Status 2026-09-26:

• PLAN: backfill wspr.bronze 2026-06 → present (~660M spots) from wspr.live #5 closed. wspr.bronze holds 2026-06 through 2026-09 (218.3M / 213.1M / 200.5M / 145.1M through the 26th), each completed month matching wspr.live.
• contest-ingest: add timestamp validation to reject typo'd dates #2 closed as superseded: typo'd dates are kept and tagged (off-declared-year 538,004 rows, timestamp-unrepresentable 75).
• internal/wspr: go test fails at setup — bridge.c in a package that does not use cgo #33: fix in Remove dead internal/wspr/bridge.c so make test can gate again (#33) #56 (internal/wspr/bridge.c was dead; make test exits 0 with it removed). Awaiting merge.
• llm@bench70.service: not broken. The benchmark instance was stopped on 09-21, and llama-server took longer than the unit's podman stop --time 30, so podman SIGKILLed it (137) and systemd recorded the stop as a failure. Clearing it needs sudo: sudo systemctl reset-failed llm@bench70.service. If a 70B instance does this on every stop, the fix is a longer stop grace in fleet-ops' llm@.service. One occurrence isn't enough evidence to change it yet.
• solar.bronze_pre_rebuild_20260922: 78,338 rows, nothing references it. Waiting on Judge's acceptance of the solar rebuild before the drop.

Also done under housekeeping: missing tags pushed (ionis-apps v4.8.0, ionis-core v4.5.0, v4.5.1, v4.5.2). VERSION files corrected: ionis-cuda#3 (3.2.0 → 4.0.2; its RPM had been building libionis-cuda.so.3.2.0) and ionis-training#2 (4.0.3 → 4.0.4).

[bob-ionis]

• solar.bronze_pre_rebuild_20260922 dropped on Judge's go, 2026-09-26 (78,338 rows, 923 KB; nothing referenced it).

[bob-ionis]

• llm@bench70.service reset by Judge. systemctl --failed on 9975 now lists 0 units. Remaining on download/ingest housekeeping: close #5 and #2, fix #33, bench70 unit, old solar backup #41: internal/wspr: go test fails at setup — bridge.c in a package that does not use cgo #33's fix (Remove dead internal/wspr/bridge.c so make test can gate again (#33) #56) awaiting merge.