feat(retain): verbatim, chunks modes and named retain strategies

[nicoloboschi]

Summary

• `verbatim` extraction mode: each chunk stored as a single memory unit with original text preserved exactly — the LLM still runs to extract entities, temporal info, and location (saves tokens by omitting the `what` field from the LLM response schema)
• `chunks` extraction mode: completely bypasses the LLM — chunks stored as-is with zero token usage; only user-provided entities are used
• Named retain strategies: define reusable sets of config overrides (e.g. `"fast"`, `"detailed"`) in bank config; select per-request or per-item via a `strategy` field
• Per-item strategy: both `MemoryItem` and `FileRetainMetadata` accept a `strategy` field that overrides the request-level strategy for that item
• Bank config UI: Retain Strategies editor and Default Strategy input added to the bank config page
• All new modes and strategy fields are configurable globally (env), per-tenant, and per-bank

How it works

verbatim mode

1. `VERBATIM_FACT_EXTRACTION_PROMPT` instructs the LLM to extract only metadata (`who`, `when`, `where`, `fact_type`, `entities`) — no `what` field in the schema
2. `_collapse_to_verbatim()` post-processing enforces 1 fact per chunk, overrides `fact_text` with the raw chunk text, merges entities from any extra LLM facts
3. Rest of the pipeline (embeddings, entity linking, semantic/temporal links) runs unchanged

chunks mode

• `_extract_facts_chunks()` returns one `ExtractedFact` per chunk using pure Python, no LLM queue, zero token usage
• User-provided entities from `RetainContent.entities` are the only entity source

Named strategies

• Defined as `retain_strategies: { name: { field: value, ... } }` in bank/tenant config
• Strategies are additive overrides on top of the resolved config — only hierarchical fields are applied
• Strategy resolution order: `request.items[i].strategy` → `request.strategy` → `retain_default_strategy` → bank/global config
• Strategy overrides can include `retain_extraction_mode`, `retain_chunk_size`, `entity_labels`, `entities_allow_free_form`, and any other hierarchical config field

Changes

• `config.py`: add `verbatim`, `chunks` to `RETAIN_EXTRACTION_MODES`; add `retain_default_strategy`, `retain_strategies` hierarchical fields
• `config_resolver.py`: add `apply_strategy()` function; validate empty strategy name keys on update
• `fact_extraction.py`: add `VerbatimExtractedFact` schema, `VERBATIM_FACT_EXTRACTION_PROMPT`, `_collapse_to_verbatim()`, `_extract_facts_chunks()`
• `memory_engine.py`: thread `strategy` param through retain pipeline and file convert handler
• `http.py`: add `strategy` to `MemoryItem`, `FileRetainMetadata`; group items by effective strategy; add `operation_ids` to `RetainResponse`
• `bank-config-view.tsx`: Retain Strategies editor with named strategy tabs, Default Strategy selector, entity fields merged into strategy form, confirmation dialog for removal
• `bank-selector.tsx`: strategy field in Add Document dialog (per-file in tabbed collapsible for file uploads)
• `route.ts` / TypeScript SDK: fix `strategy` forwarding through control plane proxy and SDK layer
• `openapi.json` + generated TypeScript client: regenerated with `strategy` fields
• `configuration.md`: document all new modes and strategies with JSON examples
• `tests/test_retain.py`: unit tests for collapse, strategy application, `chunks` mode, verbatim (LLM), per-item strategy grouping, and end-to-end integration test

Also fixes (pre-existing)

• `main.py`: missing reranker fields in CLI override block
• `cross_encoder.py`: ty type error on module attribute assignment