feat(synthesize): self-heal length-truncated batches with minimal-reasoning retry

[alpertarhan]

Implements the retry portion of #62 (option 1: self-healing retry, plus the preflight notice from option 3). Review context: #62 (comment)

Why

On local OpenAI-compatible reasoning servers, thinking tokens share the per-call max_tokens budget with the answer. When thinking runs long, the batch synthesis ends with stop reason length — a well-formed prefix, but the batch contract (one ### CHUNK n: section per chunk, terminal stop) can never be met. Until now that path went straight to the deterministic fallback, losing LLM summary quality for the whole batch.

What

Retry once at minimal reasoning (src/phases/synthesize.ts):

• length stop → one retry with reasoning: "minimal" explicitly pinned in the call opts. This caps thinking for providers that honor it, whether the configured level was high or the level was unset (provider default — the trap from the issue).
• Skipped when the configured level is already minimal (retry would be a no-op).
• Skipped for non-length format failures (duplicate ids, missing sections — those indicate a different problem).
• A failed retry still throws BatchSummaryFormatError, so the existing deterministic fallback + warning path is unchanged.

Preflight notice (src/app/steps/synthesize.ts): one info notice per run when summaryThinkingLevel is unset — the provider default may consume the batch output budget; now users know why a batch took two calls instead of guessing.

Budget interaction

A retry at most doubles one batch's wall time; with the raised watchdog ceilings from #67 (merged) the interaction is safe. Retry metrics flow through the normal trackedComplete budget accounting — two real calls are recorded as two.

Verification

• 985 tests (4 new: retry success asserting per-call reasoning opts, retry-also-fails, minimal-config skip, non-length no-retry), bench, build, release audit — all green