multi: stream live chat feedback

Add live model-stream observers so chat can render assistant and
reasoning deltas as soon as the provider emits them. The core also
reports grouped tool batches and coarse model/tool timing without
coupling terminal rendering into the turn runner.

Update the chat renderer with an animated working line, grouped
multi-tool summaries, streamed dot blocks, richer turn footers, and
less aggressive diff truncation. Mutation diffs now keep context
readable while still coloring additions and removals.

Extend status reporting with tool batch counts and approximate timing
from JSONL event gaps. Refresh README and architecture notes to
document the observer boundary and terminal behavior.

Author: bhandras

README.md

@@ -46,8 +46,9 @@ Harness currently has:
 - external process hooks for session, turn, prompt, context, tool, and
   compaction events
 - explicit config-based stdio plugin tools over a small JSONL protocol
-- human-readable tool call and tool result rendering, including colored live
-  diffs for file edits and replacements
+- streaming terminal feedback with an animated working line, grouped tool-call
+  batches, compact tool output, and colored live diffs for file edits and
+  replacements
 
 The compiled default provider is an offline echo model, so the CLI can run
 without network access. Project config can change that default. Use the
@@ -257,10 +258,11 @@ Inside chat, use slash commands for local session and context operations:
 layers, active summaries, raw replay size, and approximate token counts.
 
 `/status` reports what has already happened in the session: age, turns, model
-calls, tool calls, compactions, message bytes, and provider-reported token usage.
-When providers report usage, Harness appends `model.usage` events to the JSONL
-log and sums them for `/status`, including input, cached input, output,
-reasoning output, and total tokens when available.
+calls, tool calls, tool batches, compactions, message bytes, approximate timing
+from JSONL event gaps, and provider-reported token usage. When providers report
+usage, Harness appends `model.usage` events to the JSONL log and sums them for
+`/status`, including input, cached input, output, reasoning output, and total
+tokens when available.
 
 ## Built-In Tools