Sub-area #1 of the v0.6.6 UI redesign (issue #121).

User cells now lead with `▎` (solid bar, no animation — input is finished).
Assistant cells lead with `●`. While streaming, the bullet pulses on a
2-second cycle between 30%..100% brightness via `palette::pulse_brightness`;
once a turn completes the bullet sits at full DEEPSEEK_SKY so finished
history reads as solid.

Honors `low_motion`: the pulse is suppressed and the glyph holds full
brightness regardless of streaming state. Pager / clipboard exports
(`transcript_lines`) also skip the pulse so screenshots are stable.

Existing pager titles in `ui.rs` (`history_cell_to_text`) keep the literal
"You" / "Assistant" wording — those drive the modal title bar and read
better as words than as glyphs.

Tests: glyphs replace the literal labels in both User and Assistant cells;
streaming pulse demonstrably dips below source brightness; idle and
low_motion both pin to source.

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

Sub-area Hmbown#1 of the v0.6.6 UI redesign (issue Hmbown#121).

User cells now lead with `▎` (solid bar, no animation — input is finished).
Assistant cells lead with `●`. While streaming, the bullet pulses on a
2-second cycle between 30%..100% brightness via `palette::pulse_brightness`;
once a turn completes the bullet sits at full DEEPSEEK_SKY so finished
history reads as solid.

Honors `low_motion`: the pulse is suppressed and the glyph holds full
brightness regardless of streaming state. Pager / clipboard exports
(`transcript_lines`) also skip the pulse so screenshots are stable.

Existing pager titles in `ui.rs` (`history_cell_to_text`) keep the literal
"You" / "Assistant" wording — those drive the modal title bar and read
better as words than as glyphs.

Tests: glyphs replace the literal labels in both User and Assistant cells;
streaming pulse demonstrably dips below source brightness; idle and
low_motion both pin to source.

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

🔴 Fix Hmbown#1 (Embedder blocks async runtime):
- Pre-warm embedder in init_vector_db() via warmup_embedder()
- Model downloaded during startup, not first retrieval
- Eliminates 5-10s UI freeze on first embed call

🔴 Fix Hmbown#2 (Compaction summary pollution):
- store_compaction_summary_to_vector_db() now extracts only
  the summary section from the SystemPrompt, stripping
  boilerplate headers and workflow context
- Uses regex extraction between ## Summary and ## Workflow markers

🔴 Fix Hmbown#3 (No vector index → full table scan):
- New ensure_vector_index() in LanceDbBackend
- Checks list_indices() for existing IVF-PQ index on embedding column
- Creates missing index via table.create_index(["embedding"], Index::Auto)
- Called from ensure_tables() for all three tables

🔴 Fix Hmbown#4 (Compaction threshold regression for non-vector users):
- Added MINIMUM_AUTO_COMPACTION_TOKENS_WITHOUT_VECTOR = 500K
- should_compact() now accepts vector_db_enabled: bool
- Non-vector users keep 500K floor; vector users use 50K floor
- auto_floor_tokens=0 (explicit disable by tests) bypasses both floors
- token_threshold default restored to 800K (fallback only;
  real values come from compaction_threshold_for_model_and_effort)

🟡 Fix Hmbown#7 (No score threshold → noise injection):
- search_memories() and search_summaries() now filter
  results where score < 0.4 before returning
- Prevents low-quality retrieval from polluting system prompt

Tests: 62/62 compaction ✓, 17/17 vector_db ✓

🔴 Hmbown#1 Embedder blocks async runtime:
  - Pre-warm embedder in init_vector_db() via warmup_embedder()
  - Model downloaded at startup, eliminating 5-10s UI freeze

🔴 Hmbown#2 Compaction summary boilerplate pollution:
  - store_compaction_summary_to_vector_db() now extracts only
    the actual summary section using regex, stripping headers
    and workflow context from embeddings

🔴 Hmbown#3 No vector index → full table scan:
  - New ensure_vector_index() checks list_indices() for IVF-PQ
  - Creates missing index via create_index(["embedding"], Auto)
  - Called from ensure_tables() for all 3 tables

🔴 Hmbown#4 Compaction threshold regression:
  - Added MINIMUM_AUTO_COMPACTION_TOKENS_WITHOUT_VECTOR = 500K
  - should_compact() accepts vector_db_enabled: bool
  - Non-vector users keep 500K floor; vector users get 50K
  - auto_floor_tokens=0 bypasses both (test compatibility)
  - token_threshold default restored to 800K

🟡 Hmbown#7 Score threshold filtering added:
  - search_memories() and search_summaries() filter score < 0.4

Tests: 62/62 compaction ✓, 17/17 vector_db ✓, 99/100 engine ✓

… precision

- resolve_path: normalise .. components; reject workspace escapes (Hmbown#1,Hmbown#5)
- tool_read_file: clamp start_line >= 1 to prevent underflow panic (Hmbown#4)
- tool_edit_file: use replacen(.., 1) to avoid mass-replacement bugs (Hmbown#6,Hmbown#8)

)

When an OpenAI-compatible backend (vLLM, Ollama, LM Studio, Together AI,
self-hosted vLLM/SGLang, etc.) streams an assistant message containing
multiple tool_calls in a single round, only the **last** tool's
`Event::ToolCallStarted` was firing. The preceding N-1 tool calls
executed and produced tool_result events, but never announced their
start to consumers (TUI / runtime API / embedder bridges), leaving them
with N orphan tool_result blocks and no matching tool_use blocks in the
assistant history.

## Reproduction

```text
backend dispatches:   7 × write_file + 1 × exec_shell
log shows:            7 × ApprovalRequired events ✓
listeners receive:    1 × chat:tool_start, 7 × chat:tool_end
session history:      1 tool_use + 7 tool_result (6 orphans)
```

Tested against vLLM 0.7 + Qwen3.6-35B-A3B with a "scaffold 7-file Tauri
template" prompt. Any model+backend combo that emits batch tool_calls
trips this — typical when a single LLM round asks for multiple parallel
file writes or edits.

## Root cause

`run_turn` tracked the currently-streaming tool block with a single
`current_tool_index: Option<usize>`. The Anthropic-style adapter
(non-streaming response → events at `chat.rs::L1807`) emits
Start/Stop pairs in lockstep so the slot never overlaps. But the
OpenAI streaming parser (`chat.rs::L1954-2064`) emits every
`ContentBlockStart::ToolUse` as soon as a tool_call delta lands, then
batches every `ContentBlockStop` at `finish_reason`:

```text
Start { index: 0 }       // tool #1
Delta { index: 0, .. }
Start { index: 1 }       // tool #2 — overwrites current_tool_index
Delta { index: 1, .. }
…
Start { index: 6 }       // current_tool_index = Some(6)
Delta { index: 6, .. }
Stop  { index: 0 }       // take() returns Some(6)  ← wrong tool!
Stop  { index: 1 }       // take() returns None
Stop  { index: 2 }       // take() returns None
…
```

The first `Stop` consumes the last index and emits `ToolCallStarted`
for the wrong `tool_uses` entry; every subsequent `Stop` finds the
slot already `None` and skips the entire `if let Some(index) = …`
branch, dropping the announcement.

## Fix

Replace the single slot with `HashMap<u32 block_index, usize
tool_uses_idx>`:

- `ContentBlockStart::ToolUse` and `::ServerToolUse` insert the
  `(event.index → tool_uses.len())` mapping.
- `InputJsonDelta` looks up by the `ContentBlockDelta` outer index.
- `ContentBlockStop` removes by the stop's index, so each Stop routes
  to its own `tool_uses` entry regardless of arrival order.

Routing no longer depends on `current_block_kind` (which has the same
single-slot overwrite problem); `current_tool_indices.remove(&index)`
returning `Some(_)` already proves the Stop belongs to a tool block.

## Tests

Added `batch_tool_calls_preserve_all_tool_use_indices` in
`core/engine/turn_loop.rs::tests` — feeds 7 Starts and 7 Stops through
the same `HashMap` API used by `run_turn`, asserts every index round-trips.

Manual end-to-end verification: vLLM + Qwen3.6-35B + 7-file Tauri
template prompt → frontend `messages` history now contains all 7
`write_file` tool_use blocks paired with their tool_result blocks.

Co-authored-by: hexin <he.xin@h3c.com>

…trol, verify fast-path, naming

- combined_hash now hashes full tool JSON (name + description + schema),
  not just tool names, so schema/description changes are detected (Hmbown#1).
- build_messages preserves cache_control from SystemPrompt::Blocks so
  DeepSeek context-cache breakpoints are not lost (Hmbown#2).
- FrozenPrefix::verify() compares raw text before hashing to avoid
  redundant SHA-256 computation (Hmbown#3).
- build_messages uses self.message_count() for vector capacity (Hmbown#4).
- Strict-mode error message no longer references nonexistent
  /prefix-unfreeze command — suggests restart or config toggle (Hmbown#5).

Six fixes for the bot review comments landed on PR #2864 head
649d399. See phase2-playbook.md §7 for the triage rationale.

* persistence.rs: oversized state file now surfaces an InvalidData
  error instead of silently returning a default. The old behaviour
  would let the next save overwrite the oversized file and destroy
  the user's data. Test updated to expect the error.

* persistence.rs: PersistedDelegation gains a `status` field so
  in-flight `InProgress` delegations aren't silently demoted to
  `Pending` on restart. The snapshot now writes the live status
  and restore_from_snapshot honours it. Adds a regression test.

* mention.rs: resolve_tab_mention no longer sorts its input — tab
  mentions (@tab2) must map to the visual order in the tab bar,
  not to an arbitrary ID sort. Test updated.

* manager.rs: `pending_tasks` renamed to `completed_delegations`
  because the getter returns completed DelegationResults, not
  in-flight tasks. Docstring points to the in-flight getter
  `pending_delegations` to avoid the same confusion recurring.

* manager.rs: delegate_task and start_meeting now validate that
  the from/to tab IDs (or all participant IDs) currently exist
  in the manager. Returns `None` on any unknown ID, preventing
  orphaned tasks / meetings with stale tab references. Two new
  regression tests cover both methods.

Local CI matrix (Windows runner, flags matching ci.yml):
- cargo fmt --all -- --check: exit 0
- cargo clippy --workspace --all-features --locked -- -D warnings: exit 0
- cargo test --workspace --all-features --locked: 4282 pass, 6 fail
  (the 6 failures are pre-existing on the baseline; not caused
  by this PR)
- git diff --exit-code -- Cargo.lock: exit 0

The three deferred threads (#1 close_tab cleanup, #5 cross_tab_links
snapshot, #8 TabGroup::new collision risk) are explicitly out of
scope here; they belong to the follow-up collab/UI PR per the
narrow-harvest promise to Hmbown.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>