Addresses review feedback on the discovery hardening:
- parsePiModels: the previous guard only dropped the `Warning:`-prefixed
form (which collapses to an empty `Warning/` half). The real pattern
warning can arrive without that prefix — `No models match pattern "..."` —
and the table branch would turn it into a bogus `No/models` model. Filter
pi's diagnostic lines (unmatched-pattern message, `Warning:`/`Error:`/`Info:`
prefixes) before field-splitting, with the empty-half check kept as a
structural backstop.
- discoverOpenCodeModels: parse the verbose output even on a non-zero exit and
only fall back to plain `opencode models` when nothing parses. The earlier
version skipped the fallback whenever verbose printed any bytes, so
unparseable noise on a non-zero exit returned empty — worse than before.
Tests: pi non-zero-exit cases now assert the result is exactly the resolvable
model (no junk row) and cover the un-prefixed warning; new OpenCode test covers
verbose-noise-then-fallback.
Co-authored-by: multica-agent <github@multica.ai>
5744e080ef