docs(harness-view): add on-device LLM Chat workflow

The Workflow view's Core tab listed seven workflows but had no entry for
the chat flow that landed in v0.9.0 (Gemma 4 + Nemotron 3 Nano), so the
viewer's "how does this app actually do X" picture stopped at search /
graph / TUI / GUI. Adds an `on-device-llm-chat` entry with both
diagrams:

  * mermaid — high-level request loop: model present? → optional
    ModelDownloader → GgufReader probe → GpuEnumerator/CtxRecommender →
    LlmHost load → ChatTemplateRegistry pick → AgentChatLoop's
    JSON-tool-call cycle against CodeScanToolbelt (SqliteStore + FS),
    terminating on `tool=done`. Forensic log path called out.

  * sourceMermaid — file/class layout grouped by directory (Tui/,
    Services/Llm/, Services/Llm/Tools/, Services/) showing the wiring
    between ChatView and the toolchain. Annotates the two chat
    templates by their distinguishing turn markers so the family
    boundary is visible.

Also patches the existing `tui-flow` entry: its Home-menu diagram was
missing the Chat node entirely and its sourceMermaid only listed
TuiApp.cs, so adds the Chat node with a `see on-device-llm-chat`
back-edge to the LLamaSharp host, plus ChatView.cs and the new
StartupTracer in the source panel.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: smpark-luna

Home/harness-view/data/workflow-graph.json

@@ -25,9 +25,16 @@
     {
       "id": "tui-flow",
       "label": "Interactive TUI",
-      "description": "tui command — Terminal.Gui v2 (NetDriver) state machine. Project list → scan/search/graph screens, all reading from the same SqliteStore",
-      "mermaid": "flowchart LR\n  U[codescan tui] --> APP[TuiApp]\n  APP --> HOME[Home menu]\n  HOME --> PRJ[Projects screen]\n  HOME --> SCAN[Scan screen]\n  HOME --> KW[Keyword search]\n  HOME --> GR[Graph search]\n  PRJ --> ST[(SqliteStore)]\n  SCAN --> ST\n  KW --> ST\n  GR --> ST",
-      "sourceMermaid": "flowchart TB\n  subgraph Tui[Tui/]\n    TA[TuiApp.cs<br/>state machine root]\n  end\n  subgraph Svc[Services/]\n    ST[SqliteStore.cs]\n    DS[DirectoryScanner.cs]\n    SA[SourceAnalyzer.cs]\n  end\n  TA --> ST\n  TA --> DS\n  TA --> SA"
+      "description": "tui command — Terminal.Gui v2 (NetDriver) state machine. Home menu fans out to scan/search/graph/projects/chat screens, all sharing the same SqliteStore. Chat descends into the on-device LLM workflow",
+      "mermaid": "flowchart LR\n  U[codescan tui] --> APP[TuiApp]\n  APP --> HOME[Home menu]\n  HOME --> PRJ[Projects screen]\n  HOME --> SCAN[Scan screen]\n  HOME --> KW[Keyword search]\n  HOME --> GR[Graph search]\n  HOME --> CHAT[Chat screen]\n  PRJ --> ST[(SqliteStore)]\n  SCAN --> ST\n  KW --> ST\n  GR --> ST\n  CHAT -->|see on-device-llm-chat| LLM[(LLamaSharp host)]",
+      "sourceMermaid": "flowchart TB\n  subgraph Tui[Tui/]\n    TA[TuiApp.cs<br/>state machine root]\n    CV[ChatView.cs<br/>chat sub-state machine]\n  end\n  subgraph Svc[Services/]\n    ST[SqliteStore.cs]\n    DS[DirectoryScanner.cs]\n    SA[SourceAnalyzer.cs]\n  end\n  subgraph Diag[Services/Diagnostics/]\n    STR[StartupTracer.cs<br/>cold-start phase log]\n  end\n  TA --> ST\n  TA --> DS\n  TA --> SA\n  TA --> CV\n  TA -.marks.-> STR"
+    },
+    {
+      "id": "on-device-llm-chat",
+      "label": "On-device LLM Chat (Gemma 4 / Nemotron 3 Nano)",
+      "description": "TUI Chat — picks a GGUF (Gemma 4 E4B/E2B or NVIDIA Nemotron 3 Nano 4B), loads it through LLamaSharp on CPU or Vulkan GPU, and drives a JSON tool-call loop against the indexed codebase. The model is grammar-constrained (GBNF) to emit ONE tool-call JSON per turn; CodeScanToolbelt executes the call against SqliteStore / the local filesystem and feeds the result back as the next turn's input. Turn markers + antiprompts are family-aware — Gemma 4 (`<|turn>…<turn|>`) and Nemotron 3 Nano (ChatML + /no_think) are mutually incompatible, so ChatTemplateRegistry picks the right one from the catalog entry, the filename, or the GGUF `general.architecture` field. Runs fully offline — no network after the one-time model download. Every raw model emission and tool result is appended to `~/.codescan/logs/chat-*.log` for postmortem.",
+      "mermaid": "flowchart LR\n  U[TUI Chat screen] --> CV[ChatView]\n  CV --> SEL{GGUF on disk?}\n  SEL -->|no| DL[ModelDownloader<br/>HTTP Range resume → .part]\n  DL --> SEL\n  SEL -->|yes| GG[GgufReader<br/>arch + ctx + KV dims]\n  GG --> GE[GpuEnumerator<br/>Vulkan + nvidia-smi]\n  GE --> CR[CtxRecommender<br/>ctx fits selected device]\n  CR --> LH[LlmHost.LoadAsync<br/>CPU or Vulkan offload]\n  LH --> CTR[ChatTemplateRegistry<br/>family pick]\n  CTR --> ACL[AgentChatLoop<br/>StatelessExecutor + GBNF]\n  Q[User question] --> ACL\n  ACL -->|JSON tool call| TB[CodeScanToolbelt]\n  TB -->|db_search / project_tree / project_info / list_projects / graph_query| ST[(SqliteStore)]\n  TB -->|read_file / grep_file| FS[(Local filesystem)]\n  TB -->|tool result JSON| ACL\n  ACL -->|tool = done| ANS[Answer rendered as<br/>Gemma / Nemotron label]\n  ACL -.every turn.-> LOG[~/.codescan/logs/chat-*.log]",
+      "sourceMermaid": "flowchart TB\n  subgraph Tui[Tui/]\n    CV[ChatView.cs<br/>state machine + transcript + picker]\n  end\n  subgraph Llm[Services/Llm/]\n    CAT[ChatModelCatalog.cs<br/>Gemma 4 + Nemotron 3 Nano entries<br/>+ ChatModelFamily]\n    LOC[ModelLocator.cs<br/>~/.codescan/models/]\n    DL[ModelDownloader.cs<br/>resumable HTTPS]\n    GG[GgufReader.cs<br/>header probe pre-load]\n    GE[GpuEnumerator.cs]\n    CR[CtxRecommender.cs]\n    LH[LlmHost.cs<br/>LLamaSharp wrapper]\n  end\n  subgraph Tools[Services/Llm/Tools/]\n    IT[ChatTemplate.cs<br/>IChatTemplate + ChatHistoryEntry]\n    GCT[GemmaChatTemplate.cs<br/>&lt;|turn&gt; markers]\n    NCT[NemotronChatTemplate.cs<br/>ChatML + /no_think preamble]\n    CTR[ChatTemplateRegistry.cs<br/>catalog → filename → arch]\n    TG[CodeScanToolGrammar.cs<br/>GBNF + system prompt<br/>+ PATH COPY RULE]\n    ACL[AgentChatLoop.cs<br/>JSON tool-call loop]\n    TB[CodeScanToolbelt.cs<br/>tool dispatch +<br/>Did-you-mean basename recovery]\n    CSL[ChatSessionLogger.cs]\n  end\n  subgraph Svc[Services/]\n    ST[SqliteStore.cs]\n  end\n  CV --> CAT\n  CV --> LOC\n  CV --> DL\n  CV --> GG\n  CV --> GE\n  CV --> CR\n  CV --> LH\n  CV --> CTR\n  CV --> ACL\n  CTR --> GCT\n  CTR --> NCT\n  ACL --> IT\n  ACL --> TG\n  ACL --> TB\n  ACL --> CSL\n  TB --> ST"
     },
     {
       "id": "gui-flow",