fail hermes runs on api errors

[Svector-anu]

what

make the hermes adapter fail when the cli prints an http 4xx/5xx response while exiting zero. add and register a focused adapter regression test.

why

fork run 33150824247 looked successful at the workflow level, but the main harness output was only HTTP 400: modelCode: does not exist. hermes exited zero, so the adapter wrapped that error text in a success envelope; the scorer then repeated the same error and no health score was recorded.

the exact model source was traced to stale captured config: model.default: tencent/hy3:free. the official live nous catalog updated 2026-08-28 contains tencent/hy3 and no tencent/hy3:free. RH_MODEL_ARG=default was not passed to hermes because the adapter already suppresses that sentinel.

fix

• detect hermes stdout containing an http 4xx/5xx response
• emit the real api diagnostic on stderr
• exit nonzero instead of creating a false success envelope
• cover both normal output and the zero-exit api-error case

verification

mutation with only the detection removed:

rc=1
hermes HTTP error returned rc=0, want 1

restored:

hermes adapter error tests passed

real fork dispatch 33189422348 now fails at the run step with:

run-harness hermes failed: read-only: workspace write-locked via bwrap hermes API error: HTTP 400: modelCode: does not exist

this pr fixes false-green classification. selecting a current nous model and refreshing HERMES_AUTH remains an operator credential/configuration action; the adapter does not silently rewrite it.