[igs-rogenlo]

Problem

Summary completions containing only punctuation can pass the non-empty check and be stored as successful summaries. This occurs after removable reasoning blocks, such as <think>...</think>., and also with bare output such as ..., where the host may already have separated reasoning.

A downstream observation found a stored ... summary over substantive input after the previous, narrower fix. The raw provider envelope is unavailable, so this does not attribute the response shape to a particular provider or transport.

Change

Treat a body containing only Unicode punctuation, separators, and control characters as empty output regardless of whether this sanitizer removed a reasoning block. Reuse existing bounded retry, envelope recovery, and deterministic fallback. Preserve short CJK text, letters, numbers, and symbols without a minimum-length heuristic.

The change covers initial and retry content/envelope extraction. Metadata diagnostics remain; no summary or reasoning bodies are logged.

Validation

• The previous PR head 933e83c fails five additional regressions and passes four substantive controls. Failures cover both extraction paths on the initial/retry completion and the real compaction engine with SQLite persistence.
• The follow-up candidate passes 130 tests across summarization, summarizer-reasoning, and compaction-work-accounting. Typecheck, plugin build, and whitespace checks pass.
• The same additional regressions on unmodified latest upstream main c3745b0e6ee5e23dcce8c55cca5c3f9e8ec51d62 fail five cases and pass four controls. The latest-main candidate passes all 126 tests across the same three suites.

The completions in these tests are controlled fixtures; persistence uses real SQLite. Historical summaries have not been rewritten. Earlier live-provider checks exercised the narrower sanitizer and are not acceptance evidence for this follow-up.

Additional downstream validation (2026-11-05)

The sanitizer compiled from this follow-up was applied as the sole function change to the current customized 1.2.0 bundle (-18 bytes), preserving every other runtime byte. The full before/candidate bundles passed 26 controls, including actual compaction and file-backed SQLite: the predecessor stored ..., while the candidate retried and stored recovered substantive text.

Two bounded real-provider completions passed: one ordinary summary and one recovery completion after a deliberately injected bare-punctuation first response. The recovered 703-character summary was read back from the isolated SQLite store. This probe directly used the existing Qwen HTTP route; it does not establish host runtime.llm.complete normalization coverage.

The downstream Gateway subsequently loaded the exact compiled sanitizer after a native restart. The plugin was active, readiness returned HTTP 200, and 13,709 integrity checks had no mismatches. A read-only sample of the first 12 new natural summaries after confirmed activation contained no empty or punctuation-only bodies. Busy/CPU warnings remained and are tracked separately; healthy readiness is not full performance acceptance.

The bad first completion was injected. These observations do not claim a natural bad provider response was rejected and recovered by the live Gateway, nor that all downstream conversations or other PRs have been accepted.