[D3-01] Build explicit cross-paper synthesis over selected papers

[dkritarth]

Outcome

Answer comparative research questions across a user-selected small paper set with traceable per-paper evidence.

Dependencies

Blocked by #42. Do not start until durable-library gate closes.

Scope

Explicit paper selection, context-budget behavior, unified ACP prompt, streaming answer, and source mapping. Agent reads selected local markdown files directly. No vector DB or custom embeddings.

Acceptance criteria

• User explicitly chooses papers and can review context before send.
• Answer distinguishes each paper and cites sections/passages traceably.
• Conflicts, missing evidence, and uncertainty are stated.
• Context limits produce clear degrade/split behavior, not silent omission.
• Cross-workspace or trashed papers cannot leak into context.
• Session persists selected-paper provenance.
• At least three real papers are verified with comparative question.

Live verification

Compare three related papers on method, assumptions, data, and limitations. Manually check every factual claim/citation against sources; test removed paper, oversized context, restart, and Codex failure.

Required evidence

Focused PR uses Closes #43. Provide a written design first, failing tests/baseline, focused/full tests, typecheck, build, diff check, exact live Electron procedure, real-paper evidence, screenshots, console state, performance observations, and limitations. Independent reviewer repeats representative workflow.