docs: remove extraneous content — pipeline internals moved to architecture, changelog removed

- features.md: remove SQL impl detail from tasks section, sw.js reference from PWA section,
  and entire "LLM Chat — Internal Pipeline" section (moved to architecture.md)
- architecture.md: add "LLM Pipeline Internals" section (intent routing, tool loop, duplicate
  guards, context window, research pipeline, image cache)
- development.md: remove site-specific NFS path from custom runner instructions
- Remove changelog.md (duplicates git history)
- README.md: remove changelog link

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

docs/architecture.md

@@ -267,6 +267,52 @@ Session cookies: `HttpOnly`, `SameSite=Lax`, optionally `Secure`. Session includ
 
 See [sso-oauth.md](sso-oauth.md) for provider-specific setup instructions.
 
+## LLM Pipeline Internals
+
+### Intent Routing
+
+Before the main model runs, a lightweight intent classifier (`services/intent.py`) runs concurrently with `build_context()`. It makes a fast non-streaming call using a smaller dedicated model (`OLLAMA_INTENT_MODEL`, default `qwen2.5:7b`) to determine if the message requires a tool call.
+
+**Skip heuristic** — Intent classification is skipped entirely for short messages (≤10 words) with no action/object keywords, saving 400–800ms on conversational replies.
+
+**Prior-work fast-path** — `_PRIOR_WORK_REFS` regex detects phrases like "research you did", "note you made", "using your research" and returns no-tool immediately, preventing `search_web` from firing when the user references existing notes.
+
+If a tool is detected, the intent's one-sentence `ack` field is streamed as the first chunk (TTFT), the tool executes, then the main model generates a follow-up with the tool result. For chat-only responses the main model streams directly.
+
+### Tool Loop
+
+Multi-round tool loop (max 5 rounds). All implementations in `services/tools.py`; `execute_tool()` is the dispatcher.
+
+**Duplicate protection on `create_note` / `create_task`:**
+1. Exact title match (case-insensitive) → hard block, redirect to `update_note`
+2. Fuzzy title match (SequenceMatcher ≥ 82%; punctuation stripped before candidate search) → hard block
+3. Semantic content similarity (threshold 0.90, body ≥ 200 chars) → soft block with `requires_confirmation: true`
+
+**Project resolution** (`_resolve_project`): 4-step lookup — (1) exact DB match, (2) `query in title` substring, (3) `title in query` reverse substring, (4) SequenceMatcher ≥ 0.55.
+
+### Context Window and Summarisation
+
+`OLLAMA_NUM_CTX` (default 16384) controls the context window for all generation calls. Intent classification always uses `num_ctx=4096` to reduce VRAM pressure.
+
+History summarisation threshold: 30 messages. Keeps 8 recent messages. Summary max 400 tokens.
+
+### Web Research Pipeline
+
+`services/research.py` implements a full autonomous research pipeline:
+1. Intent model generates 5 focused sub-queries
+2. All 5 SearXNG queries run in parallel (200ms stagger to avoid rate limiter)
+3. Up to 15 unique URLs fetched in parallel
+4. Up to 12 sources passed to synthesis LLM
+5. Result saved as a note with `tags=["research"]`
+
+SearXNG tip: add the app server IP to `botdetection.ip_lists.pass_ip` in SearXNG `settings.yml` to bypass the rate limiter for trusted backend requests.
+
+### Image Cache
+
+`search_images` tool fetches images server-side via SearXNG, stores them on disk (SHA-256 dedup, content-type validation, 5 MB cap), and serves from `/api/images/<id>`. The user's browser never contacts the original image host.
+
+Config: `IMAGE_CACHE_DIR` (default `/data/images`), `IMAGE_MAX_BYTES` (default 5 MB).
+
 ## RAG Pipeline
 
 1. `semantic_search_notes()` — cosine similarity via pgvector, threshold configurable per call.