[plan-11] Observer / Summarizer Output Fidelity & Resilience — trust what the agent emits, or recover

Defect

claude-mem's quality depends on the observer/summarizer emitting truthful, parseable output, but nothing enforces either property. The observer SDK sometimes never receives tool results (only the calls) and then poisons; it poison-loops on prose like "No observations to record"; and provider history truncation (OpenRouter/Gemini) can drop the init prompt and the XML-output instructions, so the model emits non-XML and the parser silently drops the batch — observations stay at zero with no recovery and no signal. The fix is an output-fidelity contract: classify the observer's output (valid XML / idle-empty / prose / poisoned), recover by killing and respawning a poisoned session while preserving pending work, and protect the init/XML instructions from truncation.

Children

• #2758 — Observations never generate — observer SDK gets tool calls but not tool results, then poisons (v13.4.0)
• #2749 — Regression (v13.4.0, Windows): observer poison-loops on 'No observations to record' prose
• #2738 — OpenRouter/Gemini history truncation can drop the init prompt and XML-output instructions

Fix sequence

Design doc: plans/11-observer-output-fidelity.md. Ensure tool results reach the observer; classify output and treat idle-empty/prose as non-fatal; respawn poisoned sessions preserving pending work; pin init + XML-instruction messages so truncation cannot drop them.

Test matrix

Input	Required behavior
idle batch	classified idle-empty, not poison
prose "no observations"	no poison loop
truncating provider	XML instructions retained; output parses
missing tool results	detected; session recovered

Out of scope

Tool permissions of the observer (plan-05, shipped); write-path persistence (plan-09).