fix(ask): preserve multi-module findings when agents fail

[Li-Guangsen]

When a cross-module question reaches the map.md/agent.md path, failed readers
are silently discarded. If the Synthesizer then fails, only the first successful
reader's answer is returned. For example, auth, api, and db can all finish reading
source successfully, yet the caller receives only the auth findings.

This change retains every usable reader answer, with module attribution and
original references, when synthesis fails or returns blank output. A visible
notice explains that these findings have not been reconciled. Failed or blank
reader responses produce a separate Partial answer notice naming the missing
modules, and the synthesizer is told not to infer their behavior. Explicit
"(not relevant)" responses remain distinct from failures. The fallback does not
make another model call; no-result cases retain the existing legacy fallback.

Validation

• Added 13 focused regression cases for result retention, failed-reader coverage,
  blank output, irrelevant responses, grouped modules, and existing behavior.
• Linux/Python 3.12 full engine and CLI suite: 409 passed.
• Windows/Python 3.12 proposed regression tests: 13 passed.
• Repository contract check and all documented CLI entrypoint smoke checks pass.
• Updated the English, Chinese, and Spanish swarm documentation.

Scope and limitations

This targets the API-based map.md/agent.md multi-reader path. It does not alter
host-runner behavior, routing, retry policy, provider configuration, or public
CLI/MCP schemas. The fallback preserves findings rather than reconciling their
meaning, and may be longer than a normal synthesized answer. The tests exercise
orchestration offline; no paid-provider quality benchmark was run.