feat: pc oracle — full-wiki large-context Q&A, recompile, and gap detection

Motivation

The inject path answers per-prompt questions via top-K embedding lookup — good for latency, but structurally blind to cross-cutting concerns, global contradictions, and missing coverage. A complementary oracle subcommand dumps the full wiki into a single large-context model call, enabling questions and operations that require seeing everything at once.

The full wiki is ~132 guides × ~60 lines median ≈ 80–120K tokens. That fits comfortably in Claude Opus (200K), Gemini 2.0 Flash (1M), etc. No routing, no embeddings — raw context.

---

Modes

pc oracle <question>

Concatenate all wiki guides → send with question → stream answer.

pc oracle "What did I decide about the capture routing bottleneck?"
pc oracle "What's still marked as an open question across the wiki?"
pc oracle "Why did we reject a parallel second citation log?"

No embedding lookup, no routing. The model sees everything. For whole-system questions this is strictly better than inject's top-K.

pc oracle --interactive

REPL mode. Load the full wiki once as persistent context; query across turns without re-paying the context fill on every message.

pc oracle --recompile

Global-consistency maintenance pass. Send all guides in one call with a prompt like: "Read the full wiki. Identify contradictions, near-duplicates, and gaps. Output a revised guide set."

Why this beats the current wiki-doctor: the wiki-doctor detects near-duplicates via pairwise embedding cosine similarity — it can't see guide A saying X and guide B saying Y. A single large-context call sees both simultaneously.

Citation stability: restrict recompilation to citation-free guides (the definitional/structural layer, currently ~10% complete), or treat a recompile as a new version that re-anchors citations. See capture-destination-event-sourced-projections.md for why prose-keyed citation scrambling is the hard part.

pc oracle --gaps

Gap detection: "What topics does a developer need to understand this codebase that are NOT covered or underdefined here?" Structured audit version of the one-shot rebuild eval (wiki-dump-is-the-eval.md).

---

Future layer: --ground-truth flag

Once _citations.log gets author tags (planned in the event-sourced claim log redesign — see capture-destination-event-sourced-projections.md), add:

pc oracle --ground-truth <question>

This replaces the synthesized wiki with only user-origin citation entries — the purest "everything the user has ever said" corpus. SELECT text FROM citations WHERE author='user' is the query; the oracle is the consumer.

---

Implementation sketch

• Walk wiki_dir(), concatenate all *.md files
• Prompt: [WIKI CONTEXT]\n{all_guides}\n[QUESTION]\n{question}
• Send via existing provider abstraction (openrouter.rs / provider.rs), stream to stdout
• Model selection: expose --model flag; default to a large-context model from config (suggest a separate oracle_model config key distinct from capture_model)
• Interactive mode: loop over stdin lines, keeping messages history in memory

---

Relationship to existing commands

Command	Mechanism	Best for
pc query	Embedding lookup, top-K	Fast symbol/concept lookup
pc inject	Route + compile, top-K guides	Pre-prompt context injection
pc oracle	Full-wiki context dump	Whole-system Q&A, consistency, gaps

The inject and oracle paths are complementary, not competing. Inject is latency-sensitive (hot path); oracle is for deliberate, interactive use.

Start by tracing the existing CLI commands and provider abstraction in openrouter.rs and provider.rs, then inspect wiki_dir() and the cited wiki documents. Define how full-wiki loading, model selection, streaming, interactive history, recompilation, and gap detection fit together. Done means the oracle modes and --model configuration work without changing the existing query or inject paths.