fix(core): recover when the router stops a repetitive generation

[TaewoooPark]

Problem

In a long interactive session (the SP3 micromagnetics benchmark), the conversation ended twice with nothing on screen. Both times Motif-3 had been reasoning for more than 10k tokens when Infron stopped the generation for repetition. The router does this inside a 200 response: it writes an error chunk into the SSE stream, then closes with an ordinary stop.

{"choices":[],"error":{"code":"repetition_detected","message":"Repetition was detected in the model's output and generation was stopped. Retry, or send a `repetition_penalty` above 1.","type":"generation_error"}}

readStream read usage and cost from chunks without choices and dropped the rest, so the error vanished. What was left looked like a finished reply with no text and no tool call. In a chat, a reply without an action ends the task, so the task ended with no summary, and the reasoning was hidden.

Changes

• fix(core): retry a generation the router stops for repetition
  • An error chunk in the stream becomes a retryable TransportError of kind generation, with the server's code.
  • The loop samples the same request again at once, up to twice. Only the retry carries repetition_penalty: 1.05, as the router asks. The cut attempt is dropped.
  • If both retries are cut off too, the turn is handed back through the no-action path instead of ending the session.
  • Every other request is unchanged, in body and in recorded hash.
• fix(core): keep a chat task going after an empty reply
  • A turn in which the server sent no text is handed back like any turn without an action, instead of ending the conversation silently.
  • A chat UI test whose fake model replied with nothing now replies ok.

Evidence

• Frequency: 2 of 305 recorded responses were cut off, and both were over 10k completion tokens.
• The two cut-off requests were replayed against Infron, twice per setting. A tool call came back in:

setting	tool call reached
unchanged	3 of 4
repetition_penalty 1.05	4 of 4
repetition_penalty 1.1	3 of 4

• Seeds are not deterministic on this endpoint, so even a plain retry draws a new sample.

Tests

pnpm typecheck, pnpm test (1644 passed), pnpm lint:tools and pnpm build all pass. New tests cover:

• the error chunk in the stream
• the body and hash staying unchanged without a penalty
• the retry with the penalty
• the fallback when the retries are cut off too
• an empty chat reply

[TaewoooPark]

Maintainer review completed for 99daaed8e7bcc8022db1806c7a1efd92cb4d168f.

Live Motif model testing is currently unavailable because the Infron endpoint is unavailable to us. I did not repeat the historical live experiments described in this PR. The changed SSE parsing, retry request, retry limits, cancellation, empty-reply recovery, and record/replay behavior can be verified deterministically without a live model, so I have decided to merge based on those checks. This does not establish current endpoint health or that a penalty of 1.05 improves current real-model output quality.

An independent eight-case suite exercised the actual HttpTransport and runLoop with fragmented SSE byte streams: all eight passed at this head; seven failed on the unchanged baseline, while the unchanged HTTP/JSON error-compatibility case passed. Checks covered discarded partial tool calls, retry-only penalty and reset on the next turn, bounded exhaustion, abort before another request, whitespace-only responses, EOF error chunks, and recording/replay hashes. The existing 93 loop/transport/stream tests and 24 replay tests also passed, along with frozen install, typecheck, build, and tool-schema lint.

The regeneration budget is two retries per model turn; exhaustion returns through the existing bounded no-action path. All five head CI jobs passed: https://github.com/TaewoooPark/Motifcode/actions/runs/36267843721 . The current-main combination of #41, #40, and this PR also passed all 1,653 tests plus two cross-path checks combining paste normalization, generation recovery, and normal/cancelled tool batches. Python CI's five optional GPU-only skips remain outside this validation.

No blocking defect or mandatory follow-up implementation commit found. Merging as a squash commit after the preceding main checks complete. This is a source merge, not an npm release; no unrelated issue is being closed for this PR.