bug: recover context-limited length completions before durable output #44609

Problem

A provider can return reason: "length" before durable assistant output when the prompt plus the configured output budget exceeds the model context window. The session runner currently persists that length finish and stops instead of compacting the context and retrying.

Reproduction

With a model limit of context = 20_000 and output = 1_000, return a step-finish event with reason: "length", inputTokens: 19_900, and outputTokens: 40, without text, reasoning, or tool events. The runner makes one request; the expected sequence is original attempt, compaction summary, then retry.

A deterministic regression test on the unmodified source fails with:

Expected length: 3
Received length: 1

Proposed Fix

Treat that narrow event shape as recoverable only before an assistant message has started, then reuse the existing one-time overflow-compaction path. Do not retry when the response actually reached the configured output limit.

Verification

The failing-first regression and an output-limit guard test are included in the accompanying PR. The current Windows worktree cannot complete the green run because Bun reports EPERM while resolving existing .bun dependency links, before the test module reaches its assertions; the PR describes this limitation explicitly.

Related upstream precedent: earendil-works/pi#7540.