Reasoning models return content=null with reasoning_present at auto effort — E1 dead-letters (new §2.4 observation)

[fazpu]

Live evidence (glm-4.7-flash arm, 2026-07-27): every E1 ContextPrefix call died with

ContextPrefix: provider returned no completion content
(finish_reason='stop', content=null, reasoning_present=True,
 refusal_present=False, model='z-ai/glm-4.7-flash')

→ embed_chunk dead-letters → E2+ never runs → empty store. The #141 diagnosis metadata made this immediately identifiable (it is exactly what reasoning_present=True exists for).

Cause: at default/auto reasoning effort, some reasoning models put everything in the reasoning channel and leave content null. The earlier deepseek arm worked because REMEMBERSTACK_OPENROUTER_REASONING_EFFORT=none was pinned; removing that pin for the luna experiment exposed the class. A retry with effort=none is in flight to confirm the mitigation.

Suggestions:

1. Document in design §2.4 as a third verified provider-behavior class.
2. Consider a per-model effort override (global pin is blunt: luna benefits from reasoning, flash-class models need none for content-bearing calls) — fits the Per-stage model-bindings scorecard: re-pinning a cheaper model must beat the incumbent on that stage's canaries #150 per-stage bindings scorecard.
3. Adapter policy question (deliberate, not automatic): should a content=null + reasoning_present response ever fall back to parsing the reasoning channel? Current behavior (diagnosable failure) is honest; silent fallback risks §2.4-style corruption.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GKENhTLJg1HqhbdwCmmkbc