feat: add evidence-aware draft fallback

[igorrendulic]

Summary

Support drafting can now distinguish between evidence that supports a direct answer and evidence that should trigger a fallback choice. When the local corpus is missing, weak, conflicting, or account-specific details are needed, agent surfaces are instructed to offer both a cautious answer and a DM follow-up instead of turning the DM request into knowledge-base truth.

The fallback boundary is now explicit across the local CLI context, Codex/Claude/OpenAI guidance, and operator docs: DM follow-up wording can shape a selected draft, but it is not Manual Knowledge, not indexed evidence, and not a conflict resolver.

Validation

• uv run pytest -> 115 passed
• python3 -m compileall tg_support -> passed
• git diff --check -> passed
• Ruff was not run because this repo's dev extra only installs pytest, and no ruff executable is configured.

Post-Deploy Monitoring & Validation

No hosted production monitoring is required for this local-first CLI/plugin change.

Validation after install:

• Run a normal draft-context request and confirm the JSON includes context.evidence_sufficiency.
• For a well-supported query, expect direct_answer_supported: true and no fallback recommendation.
• For missing/conflicting/account-specific evidence, expect a fallback recommendation and reason codes that the agent surfaces before draft creation.

Failure signals:

• draft-context omits evidence_sufficiency.
• Agents create a DM follow-up without first showing sufficiency reasons.
• DM fallback text appears as Manual Knowledge or indexed evidence.

Owner: support operator validating the next local plugin install.

---