feat(filter): add batch filtering and confidence calibration

[foundatron]

Closes #7

Changes

tentacle/llm/prompts.py — Add FILTER_BATCH_SYSTEM and FILTER_BATCH_USER prompt templates

• FILTER_BATCH_SYSTEM: Instructs the LLM to score multiple articles, returning a JSON array of {"index": N, "relevance": 0.XX, "reasoning": "..."} with 1-based indices.
• FILTER_BATCH_USER: Formats a numbered list of title+abstract pairs ([1] Title: ... / Abstract: ...).

tentacle/llm/filter.py — Add filter_batch() function; filter_article() unchanged

• filter_batch(client, articles, *, model, threshold, batch_size=10) -> list[tuple[float, str]]: Chunks articles into batches of batch_size, sends each batch in a single LLM call, parses JSON array response.
• On JSON parse failure for entire batch: fall back to filter_article() individually for every article in that batch.
• On partial parse failure (valid JSON but missing/malformed/out-of-range entries): use parsed results for successful entries, fall back to filter_article() for missing ones.
• Returns results in input order.
• Set max_tokens proportionally (e.g., batch_size * 100).
• Log a warning on any fallback so degraded batches are visible in production logs.

tentacle/cli.py — Update _run_scan() to use filter_batch()

• Replace the per-article filter_article() loop with a filter_batch() call over new_articles, then build relevant_articles from the results using the same threshold check.

Review Findings

• Errors: 0
• Warnings: 3
• Nits: 4
• Assessment: NEEDS CHANGES

The most impactful fix is #4 (clamp relevance scores in the batch path to match filter_article behavior). #2 (token budget) is a latent reliability issue that will cause silent cost waste. #1 is worth a defensive assertion. The code is otherwise well-structured with solid test coverage and good fallback design.