Treat a failure to reach the harness before any prompt or tool call is
admitted (connecting or creating the session) as a distinct retryable
reach failure, so a transport blip at the first stage no longer ends a
mutating run; keep post-admission readback/poll/agent failures retried
only on an explicit read-only stage and never replay admitted tool work.
Preserve the true cause when the context is already done (join ctx.Err
with the typed transport error) and report a sanitized unknown provider
error name/status instead of a blanket summary.
Make the isolated format-repair deadline proportional to the stage
timeout (floored, capped and clamped to remaining minus a margin) so a
slow-but-repairable result is recovered without ever extending the stage
deadline; the bounded ladder, value guard, deny-all formatter session,
cumulative usage accounting and no-tool-replay rule are unchanged.
Update docs/model-results.md, docs/pipeline-performance-review.md and the
packaged skill references; add stage, opencode and server fixtures.
Combine the two reviewed versions of docs/pipeline-performance-review.md.
- Local (main) review: the exact 201939-byte runs.json capture, the three
final_result_validation sub-shapes, and the 5-to-4 unresolved busy-wait
correction.
- Incoming (lane-t2) review: both captures (artifacts/222/2 and
artifacts/222/4), both stat snapshots (134 succeeded / 3 active and
135 succeeded / 2 active), plus the server-side rejection, deploy-wait,
device-lock and refined improvement-ranking analysis.
Checked: no conflict markers remain anywhere in the tree and the two sides are
composed without duplicated headings or dead text. This is a documentation-only
change to a dated analysis document; the packaged templates, the README and
topic docs, and the packaged authoring skill references are not made stale by
it, so nothing further needs updating.