feat(contextual-prefix): add local ollama prefix tier (tier 2.5)

[vinsocci]

Summary

contextual-prefix.py shipped three prefix-generation tiers:

1. anthropic-api — --allow-egress gated, sends page bodies off-machine (~$12/1k docs)
2. claude-cli — --allow-egress gated, sends page bodies off-machine (free via CC subscription)
3. synthetic — on-machine, template-based, lower quality

For users who already run a local LLM, there was no way to get LLM-quality contextual prefixes without egress — a gap between "good but leaves the machine" and "private but weak."

This adds a 4th tier between claude-cli and synthetic:

tier 2.5 "ollama" — POST each chunk to a local ollama /api/chat with the page body for context. Default-on when ollama is reachable; no flag needed for the common localhost case.

Egress posture (consistent with the existing model)

• Local ollama (127.0.0.1 / localhost / ::1) needs no flag.
• Non-localhost OLLAMA_URL requires --allow-remote-ollama, mirroring the scripts/tiling-check.py:351 default-deny precedent.
• --no-ollama skips the tier even when reachable.
• Asymmetric fallback: on ollama failure the tier drops to synthetic and does not climb to claude-cli/anthropic-api — egress was never consented to on this path, so silently egressing on local-LLM failure would violate the user's posture.

API surface

• New flags: --ollama-model (default qwen2.5:7b-instruct), --allow-remote-ollama, --no-ollama
• New functions: ollama_url / ollama_is_local / ollama_reachable / ollama_prefix
• pick_prefix_tier() and generate_prefix() extended; the anthropic-api and claude-cli paths are unchanged.

Test plan

Verified on macOS (Obsidian 1.12.7, ollama + qwen2.5:7b-instruct), all against this repo's own public demo vault:

Tier picker (49 demo pages, --peek):

invocation	result
default (local ollama up)	all 49 → tier=ollama
--no-ollama	all 49 → tier=synthetic

Example prefixes (generated by this branch, tier=ollama):

• wiki/concepts/Search Experience Optimization.md → "This chunk outlines the methodology, process, and key innovation of SXO in SEO analysis."
• wiki/concepts/Pro Hub Challenge.md → "This chunk outlines the Pro Hub Challenge, detailing its structure and rules for both challenges."
• wiki/concepts/cherry-picks.md → "This chunk outlines Tier 1 features for quick implementation in the Cherry-Picks feature backlog."

Production scale: 895 chunks across 401 pages generated with tier=ollama, zero egress, no errors.

Notes

• Pairs naturally with the existing ollama rerank stage (rerank.py), so the full contextual-retrieval pipeline (prefix → BM25 → cosine rerank) can run entirely on-machine.
• Default model qwen2.5:7b-instruct is a suggestion, not a requirement — --ollama-model accepts any pulled tag (tested also with qwen2.5-coder:14b).

🤖 Generated with Claude Code